Patent Application
20240037915
This disclosure relates generally to the optimization of preprocessing of streaming video data and, more specifically, to the optimization of preprocessing parameters to improve a main output of a main artificial intelligence model.
Cameras are beneficial for use in many areas of commercial and personal practice. For example, security cameras are used within (and outside) commercial warehouses and on private personal property. Other applications use cameras along assembly lines for quality control purposes. With the increased capabilities of cameras having higher quality imagery (i.e., resolution) and a wider field of view, more area can be shown in the streaming video by the camera. A large portion of the frame/field of view may be of little or no interest to the consumer (e.g., a security or manufacturing company). However, current practices relay the entirety of the streaming video (i.e., the entire frame/field of view) to the consumer, which can be time and resource consuming due to the need to transfer large frame (i.e., field of view), high resolution video data.
A system and method for selection/extraction, preprocessing, and publishing of video data of a region of interest (i.e., a scene) that is a subset of a field of view of streaming video is disclosed herein. The system and method can also include processing the video data by a consumer/subscriber after the video data has been published.
Streaming video data is received from a camera with a first field of view. The video data is then preprocessed, by a computer processor such as a gateway or digital container, according to preprocessing parameters defined within a runtime configuration file that is pushed down to the computer processor. The runtime configuration file can be stored and/or edited distant from the computer processor, and any edits/revisions to the runtime configuration file can be pushed to and applied by the computer processor to the streaming video data in real time to alter the preprocessing applied to the video data. The preprocessing can include formatting/cropping the streaming video data received from the camera so as to be first video data of a first region of interest (i.e., a scene) having a second field of view that is less than (shows less area than) the first field of view shown by the entirety of the streaming video data from the camera. The preprocessing as defined by the preprocessing parameters in the runtime configuration file can also include altering the first video data's grayscale, contrast, brightness, color threshold, size, blur, hue saturation value (HSV), sharpen, erosion, dilation, Laplacian image processing, Sobel image processing, pyramid up, and pyramid down (among others). The first video data can then be published to an endpoint (such as a topic on an asynchronous messaging library like ZeroMQ) for subscription and use by a first subscriber/consumer.
The first video data can then be viewed, used, and/or processed by the first subscriber. The preprocessing as defined in the runtime configuration file can be tailored to the subscriber and the needs/uses of the subscriber and the processing to be performed by the subscriber. For example, the processing performed by the subscriber after publishing of the first video data may be using an artificial intelligence (AI) model to analyze scenarios occurring on/in the first video data. The AI model may require the first video data to be in a particular size, format, etc., which can be selected and applied during the preprocessing as set out in the runtime configuration file so that the subscriber does not need to perform this preprocessing before applying the AI model. The processing, by a computer processor, of the first video data by the subscriber can be performed distant from the camera, the location at which the runtime configuration file is stored and/or edited, and the gateway/container upon which the preprocessing is performed. The first subscriber can perform the processing of the video data to determine at least one output, with the output being indicative of an inference dependent on the first video data. For example, the first video data can be processed by an AI model to determine the amount of a particular product that has passed by on an assembly line (i.e., the amount of the product being an inference dependent on the first video data). The processing can include other operations, such as applying optical character recognition, clipping the first video data to make a video having a specific duration, and/or capturing one frame from the first video data to create a static image of a specific moment of the first video data.
The preprocessing and processing system can include additional features, such as a testing system/pipeline designed to optimize/improve the preprocessing parameters in the configuration file to improve the analysis (e.g., increase identification of inferences, increase accuracy, and/or decrease latency) of an AI model used by a subscriber during the processing stage. The testing system includes a test processing pipeline (as opposed to the main processing pipeline) that receives streaming video data from the camera and alters the preprocessing parameters in a test configuration file used to preprocess the streaming video data to create test video data. The test video data is then published to a test unified endpoint for access and processing by a test AI model. The test AI model determines a test output indicative of a test inference dependent upon the preprocessed test video data.
A machine learning model uses the test preprocessing parameters defined within the test configuration file and the test output for training of the machine learning model (e.g., a training module) by determining if the test preprocessing parameters are predictive to produce a test output that satisfies the baseline criterion. The test configuration file (with the test preprocessing parameters) and the corresponding test output are saved to a database, which can include test preprocessing parameters and test output data from other iterations performed by the testing system/pipeline. The machine learning model uses the information in the database to train an implementation module using the training module. The training module determines the preprocessing parameters that are most predictive of outputs. The training module then conveys that information to the implementation module, which uses the information to determine test preprocessing parameters that result in a test AI model determining a test output that satisfies the baseline criterion. The training module continually trains with data from the database to refine the machine learning model to better predict the optimal preprocessing parameters to produce improved outputs. The implementation module operates in conjunction with the main system/pipeline. The implementation module receives incoming video data and alters the test preprocessing parameters to continually result in test outputs that satisfy the baseline criterion. The implementation module updates the main configuration file in the main system/pipeline to be similar to the optimal/improved test preprocessing parameters of the test configuration file such that main AI model determines an optimal main output that is an improvement upon previous main outputs. Each time a testing loop is performed by the testing system/pipeline (e.g., each time the test configuration file with test preprocessing parameters are applied to create test video data and the test video data is processed by the test AI model to determine the test output), the particular test preprocessing parameters and the corresponding test output are added to the database and used as an additional dataset to improve the machine learning model.
The machine learning model improves the test output, which has dependent variables that are to be improved and are at least one of the following: correctly identified inferences by the test AI model, accuracy of the test AI model, and latency of the test AI model. To improve the test output, the machine learning model determines the test preprocessing parameters (e.g., the independent variables) that edit the incoming video data that is eventually processed by the AI model to determine test output (e.g., the dependent variables). The test preprocessing parameters, which include video edits performed on the incoming video data, have independent variables that are at least one of the following: the video edits, a value of each of the video edits, and an order in which the video edits are performed on the incoming video data. The video edits applied to the incoming video data can include crop, grayscale, contrast, brightness, color threshold, resize, blur, hue saturation value, sharpen, erosion, dilation, Laplacian image processing, Sobel image processing, pyramid up, and pyramid down. The independent variables are altered by the machine learning model and their influence on the dependent variables is tracked and used to identify the independent variables that are most influential/predictive of a change in the dependent variables.
The machine learning model analyzes the test output and potentially compares the outputs to a baseline criterion. The baseline criterion to which the test output can be compared can be a value representative of correctly identified inferences, a value representative of accuracy, and/or a value representative of latency. Those values can be set to be values dependent upon the main output determined by the main AI model such that the baseline criterion is the values of the main output. Alternatively, the baseline criterion can be set at other values as determined by a user or by the machine learning model. If the test output satisfies the baseline criterion (e.g., is greater than the baseline criterion), the testing system can alter the main preprocessing parameters in the main configuration file in the main processing pipeline to be similar to the preprocessing parameters in the test configuration file. The newly altered preprocessing parameters (in the newly altered main configuration file) can be applied/performed on the incoming video data by the gateway to create video data that is then processed by main IA model to determine main outputs, with the main outputs being similar to the test outputs. If the test output does not satisfy the baseline criterion (e.g., is less than the baseline criterion), machine learning model can alter the test preprocessing parameters in the test configuration file and the testing loop can be repeated until the test output satisfies the baseline criterion.
The machine learning model can alter any video edits of the test preprocessing parameters defined within the test configuration file. The alterations can be a change in which video edits are applied/performed, an order in which the video edits are to be performed, and/or the values of the video edits. The starting values/configuration of the test video edits of the test preprocessing parameters (e.g., the independent variables) of the machine learning model can be determined by a user, from a look-up table, dependent upon previous starting values, or from other methods.
One embodiment of a method of improving a main output of a main AI model processing first video data using a machine learning model [I will add in claim language once it is finalized]
Another embodiment of a method for training a machine learning model [I will add in claim language once it is finalized].
While the above-identified figures set forth one or more embodiments of the present disclosure, other embodiments are also contemplated, as noted in the discussion. In all cases, this disclosure presents the invention by way of representation and not limitation. It should be understood that numerous other modifications and embodiments can be devised by those skilled in the art, which fall within the scope and spirit of the principles of the invention. The figures may not be drawn to scale, and applications and embodiments of the present invention may include features and components not specifically shown in the drawings.
System 10 can include machine-readable storage media. In some examples, a machine-readable storage medium can include a non-transitory medium. The term “non-transitory” can indicate that the storage medium is not embodied in a carrier wave or a propagated signal. In certain examples, a non-transitory storage medium can store data that can, over time, change (e.g., in RAM or cache). In some examples, storage media can be entirely or in part a temporary memory, meaning that a primary purpose storage media is not long-term storage. Storage media, in some examples, is described as volatile memory, meaning that the memory, does not maintain stored contents when power to system 10 (or the component(s) where storage media are located) is turned off. Examples of volatile memories can include random access memories (RAM), dynamic random-access memories (DRAM), static random-access memories (SRAM), and other forms of volatile memories. In some examples, storage media can also include one or more machine-readable storage media. Storage media can be configured to store larger amounts of information than volatile memory. Storage media can further be configured for long-term storage of information. In some examples, storage media include non-volatile storage elements. Examples of such non-volatile storage elements can include magnetic hard discs, optical discs, flash memories and other forms of solid-state memory, or forms of electrically programmable memories (EPROM) or electrically erasable and programmable (EEPROM) memories. Most generally, storage media is machine-readable data storage capable of housing stored data from a stored data archive.
System 10 can also include one or multiple computer/data processors. In general, the computer/data processors can include any or more than one of a processor, a microprocessor, a controller, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other equivalent discrete or integrated logic circuitry. System 10 can include other components not expressly disclosed herein but that are suitable for performing the functions of system 10 and associated methods of preprocessing and processing video data and other forms of data. For example, system 10 can include communication software and/or hardware for pushing/sending configuration file 14 from storage media at a location distant from gateway/container 16, such as cloud storage, to gateway/container 16 for execution/implementation on streaming video data 20.
Camera 12 can be any device capable of collecting streaming video data 20, such as a Real Time Streaming Protocol (RTSP) camera or a USB camera. Streaming video data 20 can be video data that is continuously captured/recorded by camera 12 in any suitable format. Camera 12 can be positioned/located to provide streaming video data 20 displaying entire frame 22 with a first field of view. The first field of view shown/displayed by camera 12 can be a wide field of view that shows multiple regions of interest. Video data 20 being collected, for example, can show a wide field of view of a warehouse for storing commercial products and/or an assembly line producing commercial products of which each individual subscriber 26A-26D may only be interested in a region/scene 24A-24C that is a subset of entire frame 22. Camera 12 can collect and transfer streaming video data 20 in any resolution/video quality and any format, including (but not limited to) MP4, AVI, FLV, WMV, MOV, MPEG, Motion JPEG, AVCHD, WebM, and/or MKV. Camera 12 can transfer/send streaming video data 20 to gateway/container 16 over any suitable means, including via the internet, short-range wireless technology, or any other type of wired and/or wireless connection.
Configuration file 14 is an editable file that contains preprocessing parameters that define, among other instructions, how streaming video data 20 is to be preprocessed by gateway/container 16 to create video data 21A-21C. Configuration file 14 can include numerous other instructions for gateway/container 16, including which camera 12 to connect to (i.e., receive streaming video data 20 from), what portion of entire frame 22 to select/extract to create scenes 24A-24C (i.e., how to crop frame 22 to create scenes 24A-24C), and at which endpoint 18 to publish the preprocessed scenes 24A-24C. This is discussed in detail with regards to
Gateway/container 16 can include a computer processor capable of performing instructions provided by configuration file 14, which can include preprocessing parameters that are to be applied to streaming video data 20. Gateway/container 16 can be a gateway node, edge device, container, virtual machine, or other software and/or hardware able to accept configuration file 14 and perform the instructions therein to apply the preprocessing parameters to streaming video data 20. Further, gateway/container 16 can be within a single computer hardware set up due to virtualization. Gateway/container 16 can include one or multiple storage media for storing information, such as the preprocessing parameters pushed/sent to gateway/container 16 by/from configuration file 14 and/or other information like streaming video data 20. Gateway/container 16 can be located at various locations, including adjacent to camera 12 and/or on the same network as camera 12, distant from camera 12 with streaming video data 20 being received by gateway/container 16 from camera 12 via a wired or wireless connection, in the cloud, or at multiple locations. Gateway/container 16 is in communication with configuration file 14 to accept instructions for applying preprocessing parameters. Additionally, gateway/container 16 may be configured to contact configuration file 14 to determine if configuration file 14 has been edited. If edited, gateway/container 16 can perform preprocessing (on streaming video data 20 being received) according to the newly edited configuration file 14. In other examples, gateway/container 16 can utilize preprocessing parameters or other information included within configuration file 14 on a period basis, and can utilize changes (or edits) made to configuration file 14 when encountered in a next iteration of the period.
For clarity, streaming video data 20 is unaltered video data that is received by gateway/container 16 from camera 12, whereas first video data 21A is video data that has been preprocessed by gateway/container 16 (according to preprocessing parameters defined in configuration file 14) to create scene 24A. Similarly, second video data 21B is video data that has been preprocessed to create scene 24B, and third video data 21C is video data that has been preprocessed to create scene 24C. For each of first, second, and third video data 21A-21C, the originating video data is streaming video data 20 (which can be converted to raw video data 20A as described with regards to
Gateway/container 16 can be in communication with endpoint 18 to which gateway/container 16 publishes the preprocessed video data 21A-21C (e.g., scenes 24A-24C). The communication can be wired or wireless, such as communication via the internet. However, endpoint 18 can be at the same location as gateway/container 16 or on the same computer hardware set up and/or network. Further, endpoint 18 can be located on the internet with a unique address and/or security protocol that allows for subscription and access to scenes 24A, 24B, and 24C. Scenes 24A, 24B, and 24C can be published to endpoint 18 using an asynchronous messaging library, for example ZeroMQ, such that scenes 24A, 24B, and 24C are published as topic 1, topic 2, and topic 3, respectively. Subscribers 26A-26D can subscribe to any of topics 1-3 to receive video data 21A-21C of scenes 24A, 24B, 24C, respectively. Gateway/container 16 can publish video data 21A-21C of each of scenes 24A-24C to endpoint 18 in any format suitable for use by subscribers 26A-26D. For example, video data 21A-21C can each be published as Motion JPEG or any of the formats listed above with regards to streaming video data 20. The format that video data 21A-21C of each of scenes 24A-24C can be designated in configuration file 14 and applied to video data 21A-21C by gateway/container 16.
Each scene 24A-24C at topics 1-3, respectively, can be subscribed to by any number of subscribers 26A-26D. In the example shown in
Configuration file 14 can be edited and pushed/conveyed to gateway/container 16 in real time (i.e., runtime) such that preprocessing parameters 30A-30C (and the other information contained in configuration file 14) can be applied to streaming video data 20 immediately to preprocess and output video data 21A-21C. Configuration file 14 can be stored, edited, and/or pushed/conveyed to gateway/container 16 in any suitable format/file type, such as a text file, a comma separated value file, or other format/file type. Configuration file 14 can include other information/parameters not expressly disclosed herein and not shown in
Configuration file 14 can include camera credentials 28, which provides the information needed for gateway/container 16 to connect to camera 12 and/or receive streaming video data 20 from camera 12. Camera credentials 28 can include other information such as encryption/decryption information, security access information, and/or instructions for beginning and/or ending the collection of streaming video data 20 by camera 12. Camera credentials 28 can include information for connecting to multiple cameras 12 and/or information for gateway/container 16 to receive the same or different streaming video data 20 from the same or different cameras 12 for different scenes 24A-24C. In one example, camera credentials 28 are provided once and applied to all scenes 24A-24C. In another example, different camera credentials 28 are provided for each scene 24A-24C and applied to each of scenes 24A-24C individually.
Configuration file 14 also includes information specific to each scene 24A-24C. The information/instructions are designated as preprocessing parameters 30A, 30B, and 30C, respectively. Preprocessing parameters 30A-30C are used by gateway/container 16 (e.g., in the form of executable instructions or indications of the executable instructions) and applied to streaming video data 20 to create video data 21A-21C of scenes 24A-24C, respectively. Preprocessing parameters 30A-30C can include topic name/publishing location 32, which designates where video data 21A-21C of scenes 24A-24C will be published after preprocessing. As described below with regards to gateway/container 16 and
Configuration file 14 can designate video format 34 that each of scenes 24A-24C is to be published at, which can be the same format as streaming video data 20 or any other type of suitable video format, including the formats listed above with regards to streaming video data 20 and/or video data 21A-21C. Scenes 24A-24C can be published having the same video format 34 or different video formats 34. If the format of the video data of scenes 24A-24C is to be changed, the steps of changing the video format can be performed before, during, or after any of the other instructions/steps set out in preprocessing parameters 30A-30C. For example, video format 34 can be changed before, during, or after video edits 38A-380 are performed by gateway/container 16.
Preprocessing parameters 30A-30C can also include accompanying information 36, which is information provided/published with video data 21A-21C for each of scenes 24A-24C. Accompanying information 36 can include any information about first video data 21 that may be of use to subscribers 26A-26C. For example, accompanying information 36 can include first video data 21A frame size, which may be helpful in indicating to subscriber 26A what processing should be performed on first video data 21A of scene 24A; if the frame size is 720 pixels by 486 lines, first video data 21A of scene 24A may be most suitable for processing by an AI model. Accompanying information 36 can include metadata and/or other information regarding what preprocessing has been performed on streaming video data 20 to create video data 21A-21C for scenes 24A-24C, respectively.
Preprocessing parameters 30A-30C can also include preprocessing pipeline 38 that includes numerous video edits 38A-380 that can be applied to streaming video data 20 to create video data 21A-21C for each of scenes 24A-24C. Preprocessing pipeline 38 can designate the instructions for the entire video edits made to streaming video data 20 for each of scenes 24A-24C. The amount of time gateway/container 16 takes to perform the video edits designated by preprocessing pipeline 38 (i.e., video edits 38A-380) for each of scenes 24A-24C can be measured, recorded, and displayed as preprocessing performance metrics (see
Video edits 38A-380 are a nonexclusive list of edits that can be designated in configuration file 14 and performed on streaming video data 20 by gateway/container 16. Preprocessing pipeline 38 can include other video edits not expressly included in the list of video edits 38A-380. Similarly, not all of video edits 38A-380 need to be performed to create video data 21A-21C of each of scenes 24A-24C, and different scenes 24A-24C can include different video edits 38A-380 performed on streaming video data 20 by gateway/container 16. In one example, only crop 38A is performed on streaming video data 20 to create first video data 21A of scene 24A, while a different crop 38A and brightness 38D, resize 38E, and dilation 38K are performed on streaming video data 20 to create second video data 21B of scene 24B that is different than first video data 21A of scene 24A.
Each of video edits 38A-380 are briefly described as follows. Crop 38A is the removal of unnecessary areas/regions (i.e., regions that are not of-interest to the subscriber) of entire frame 22 having first field of view to create scenes 24A-24C each with second, third, and fourth field of views, respectively. Scenes 24A-24C that have been cropped 38A have fields of view that are a subset of (i.e., less than) first field of view of entire frame 22. Grayscale 38B is the alteration of the color of video data 21A-21C and can include limiting the color to be between white and black. Contrast 38C is the alteration of the difference between the maximum and minimum pixel intensity. Brightness 38D is the alteration in the luminous brightness of video data 21A-21C. Threshold 38E is the alteration of the color of video data 21A-21C by changing the color of select pixels of video data 21A-21C that are above a specified threshold color value. Resize 38F is the alteration of the frame size of video data 21A-21C without cutting/cropping any of the frame out. Blur 38G is the alteration of the clarity of video data 21A-21C, which may be desired for some processing applications, such as an AI model, performed by subscribers 26A-26D. Hue saturation value (HSV) 38H is assigning a numerical readout of video data 21A-21C that corresponds to the color contained therein. Sharpen 381 is altering video data 21A-21C to make the objects therein appear more defined/sharpened. Erosion 38J is altering video data 21A-21C by shrinking pixels and/or removing pixels on object boundaries, while dilution 38K is the reverse of erosion in that video data 21A-21C is enlarged by resizing pixels and/or adding pixels at object boundaries. Laplacian image processing 38L and Sobel image processing 38M are processing techniques known in the art that can be applied to video data 21A-21C. Pyramid up 38N and pyramid down 380 are altering video data 21A-21C by smoothing and/or subsampling as known in the art. Each of scenes 24A-24C can include these and other video edits 38A-380 to be applied by gateway/container 16 to preprocess streaming video data 20 and output as scenes 24A-24C for use by subscribers 26A-26D.
Configuration file 14 can arrange the instructions of camera credentials 28 and preprocessing parameters 30A-30C to be performed in any order, or gateway/container 16 can have the capabilities to arrange/rearrange the information/instructions to be performed in a desired/optimized sequence. Additionally, gateway/container 16 can be configured to perform each set of preprocessing parameters 30A-30C in parallel such that preprocessing parameters 30A, preprocessing parameters 30B, and preprocessing parameters 30C are performed at the same time (and the time gateway/container 16 takes to perform those preprocessing parameters are measured, recorded, and displayed as metrics). Configuration file 14 can be edited at any time by an editor and then pushed/conveyed/accessed by gateway/container 16 at runtime such that the preprocessing of video data 21A-21C is altered according to the newly edited configuration file 14 at runtime.
Inputs to gateway 16 can be streaming video data 20 (received from camera 12) and configuration file 14, which includes camera credentials 28 and preprocessing parameters 30A-30C. Outputs from gateway 16 can be scenes 24A-24C to unified endpoint 18, which is the location at which scenes 24A-24C are published as topics 1-3, and metrics 40, which is the location at which preprocessing pipeline 38 information (i.e., the amount of time gateway 16 takes to apply preprocessing parameters 30A-30C to streaming video data 20 to create scenes 24A-24C) is published/accessible. Gateway 16 can include raw video data 20A, which is streaming video data 20 that has been published at internal topic 42 and to which gateway 16 subscribes to receive video data used to create each of scenes 24A-24C. Gateway 16 can include preprocessing pipeline optimization and preprocessing 44, which uses preprocessing parameters 30A-30C as defined in configuration file 14 to preprocess streaming video data 20 (accessed as raw video data 20A at internal topic 42) to create scenes 24A-24C. Gateway 16 publishes scenes 24A-24C to proxy location 46 at topics 1-3, respectively. Gateway 16 can then publish/relay scenes 24A-24C (having video data 21A-21C) from proxy location 46 to unified endpoint 18.
Gateway 16 receives streaming video data 20 from camera 12 and can publish the video data as raw video data 20A at internal topic 42. This configuration provides for a constant, known location of published raw video data 20A independent of where the original streaming video data 20 is received from. Thus, if the location where streaming video data 20 is being received from changes (e.g., if one camera is disconnected but another camera comes online), raw video 20A will still be accessible at internal topic 42 without the need to change where gateway 16 is looking for video data to create scenes 24A-24C, thus ensuring a smooth transition of incoming streaming video data 20. Raw video data 20A published at internal topic 42 can be configured such that only gateway 16 has access.
Gateway 16 can be provided with the information in configuration file 14 (i.e., camera credentials 28 and preprocessing parameters 30A-30C) via a variety of avenues. In one example, gateway 16 has location information of configuration file 14 and actively accesses configuration file 14. In another example, configuration file 14 is pushed/conveyed to gateway 16 once, periodically, or continuously and gateway 16 passively waits to begin preprocessing streaming video data 20 until configuration file 14 has been received. Another example can be a combination of the two above examples in that gateway 16 actively accesses configuration file 14 at the beginning of preprocessing (and continues preprocessing following those known instructions) and configuration file 14 is pushed/conveyed to gateway 16 only after configuration file 14 has been edited/revised.
Gateway 16 can perform pipeline optimization on preprocessing parameters 30A-30C As discussed above, pipeline optimization can be performed by gateway 16 (or another component) to reduce the time gateway 16 takes to preprocess raw video data 20A to create video data 21A-21C (and attach any additional information) of scenes 24A-24C and/or to increase accuracy of the preprocessing operations. Pipeline optimization can include arranging/rearranging the order in which video edits 38A-380 are performed by gateway 16.
Gateway 16 can then, according to preprocessing parameters 30A-30C defined within configuration file 14, preprocess 44 raw video data 20A (which is derived from and similar to streaming video data 20) to edit raw video data 20A to create video data 21A-21C of scenes 24A-24C. The preprocessing to create each of scenes 24A-24C can be performed individually for each scene and can be performed in parallel (i.e., simultaneously). The preprocessing performed by gateway 16 can edit each of scenes 24A-24C to the desires/needs of subscribers 26A-26C, respectively. For example, gateway 16 can preprocess raw video data 20A to crop a first field of view of entire frame 22 to eliminate areas/regions of the first field of view that are of no interest to subscriber 26A and keep a second field of view of, for example, scene 24A that is a subset of the first field of view. Thus, further processing by subscriber 26A (or subscriber 26B) does not need to be performed on the areas (regions not of-interest) eliminated/trimmed by the cropping performed during preprocessing. In this example, the cropping reduces the processing time and resources needed by subscriber 26A. The alteration of video format can be included in the preprocessing and/or can be performed before, during, or after the application of other preprocessing parameters 30A-30C.
Scenes 24A-24C, which are made up of raw video data 20A that has been preprocessed according to configuration file 14 to create video data 21A-21C, are published/sent to proxy location 44. Because scenes 24A-24C are continuous video data, scenes 24A-24C are continuously published (e.g., made available) to subscribers 26A-26C, respectively. Scenes 24A-24C are published/sent first to proxy location 44 and then relayed to unified endpoint 18. Having scenes 24A-24C first being published/sent to proxy location 44 ensures that, no matter what path scenes 24A-24C take, video data 21A-21C of scenes 24A-24C will always end up at proxy location 44 and then be relayed to unified endpoint 18. Because scenes 24A-24C always end up at proxy location 44, unified endpoint 18 always knows the location to access scenes 24A-24C and can actively look to proxy location 44 to obtain scenes 24A-24C or passively wait for scenes 24A-24C to be published/sent to unified endpoint 18.
During preprocessing to create scenes 24A-24C, the amount of time gateway 16 takes to apply all of preprocessing parameters 30A-30C to raw video data 20A to create scenes 24A-24C, respectively, (or the amount of time gateway 16 takes to apply only video edits 38A-380 of preprocessing pipeline 38, depending on the desired measurement) is measured, recorded, and published at metrics 40 for viewing by an editor or any party with access. Metrics 40 can be published on an asynchronous messaging library like ZeroMQ (similar to the publishing of scenes 24A-24C) or can be displayed on a user interface similar to
As shown in
For example, preprocessing performance metrics 40A in
Scenes 24A-24C can be published from proxy location 44 to unified endpoint 18 as topics 1-3 within an asynchronous messaging library, such as ZeroMQ. Unified endpoint 18 allows for a known, constant location to which subscribers 26A-26C can look to access scenes 24A-24C, respectively. If more scenes are created, those scenes would be published and accessible at unified endpoint 18, so subscribers 26A-26D and/or other subscribers would know where to look to access the additional scenes.
After being published/sent to endpoint 18, scenes 24A-24C are available to be subscribed to and accessed by subscribers 26A-26D. Each scene can be subscribed to by any number of subscribers as is necessary/desirable. For example, a scene may be subscribed to by numerous subscribers each running processing 48 that includes different AI models. In this example, one AI model can be determining the amount of a first product that is passing through the scene on an assembly line, while a second AI model can be determining the amount of a second product that is passing through the scene on the same assembly line. In this case, the scene is unchanged between the two AI models (i.e., between the two subscribers) but the processing performed by each subscriber after the scene has been published is different. It should be noted that one entity (e.g., a person, company, quality control sector) can subscribe to a scene multiple times and thus constitute multiple subscribers. As discussed above, the preprocessing performed to create/alter each scene can be tailored to the needs of the subscriber(s) to, for example, reduce processing 48 resources and time needed to determine at least one output that is indicative of an inference the subscriber is aiming to ascertain.
The example in
The disclosed potential instructions/edits that subscriber 26A can perform in processing 48 are as follows. AI model 48A can be a program/model that may have machine learning and can use scene 24A to determine at least one output indicative of an inference dependent upon scene 24A. The inference, for example, can be the amount of a specific product that is viewable in scene 24A over a defined period of time. AI model 48A can also be, for example, a program/model that determines how many people appear in scene 24A over a defined period of time. AI model 48A can include other capabilities and/or configurations. OCR 48B can be a program (or other configuration) that recognizes and records any characters (i.e., text) that appear in scene 24A. For example, scene 24A can be video data of a street and OCR 48B will recognize and record any text that appears on the side of a vehicle, such as a delivery truck, that is in scene 24A. Video clipping 48C can clip the video data shown in scene 24A to create a clip of a defined period of time, and/or video clipping 48C can clip the video data shown in scene 24A to create a static image of a defined moment in time. Further formatting 48D can be video edits, such as video edits 38A-380 in configuration file 14, or any other video or file formatting that are performed by subscriber 26A. For example, further formatting 48D can include cropping scene 24A to be a subset of the second field of view shown in scene 24A. Display 48E can be making scene 24A viewable on a screen or other visual display. Display 48E can also include any video formatting/reconfiguring that is necessary to effectuate the display of scene 24A. While video edits 38A-380 and further formatting 48D to create and/or process video data 21A-21C have included only edits to the video/image, edits can include editing the audio or other aspects of the video data.
The potential instructions/edits (i.e., processing 48) can be performed in parallel or series. Further, processing 48 can be configured such that instructions/edits 48A-48E work together such that one instruction/edit is prompted by an inference from another instruction/edit. For example, video clipping 48C can be configured to work in tandem with another edit/process; if AI model 48A determines that a product is defective, video clipping 48C can be prompted to record and clip a particular duration (or moment to create a static image) of scene 24A showing the defective product and save the clip (or image) for proof/validation.
System 10, with associated methods, for selection/extraction, preprocessing, and publishing of streaming video data 20 into scenes 24A-24C and for processing scenes 24A-24C is used to determine at least one output that is indicative of an inference dependent upon video data 21A-21C. System 10 reduces the preprocessing/processing time and resources necessary for accomplishing the desired output/determining the selected inference. System 10 allows for a streamlined process that extends from collecting streaming video data 20 from camera 12 to making preprocessing scenes 24A-24C available to subscribers 26A-26C for further processing and analysis/determinations. System 10 also allows for measurement, recordation, and viewing of preprocessing performance metrics 40 and optimization of preprocessing pipeline 38 (and/or preprocessing parameters 30A-30C) to reduce the amount of time and resources needed to apply preprocessing parameters 30A-30C to streaming video data 20 (i.e., raw video data 20A) to create video data 21A-21C of scenes 24A-24C.
Main system/pipeline 10 is similar in configuration and capabilities as the system described with regards to
Testing system/pipeline 310 can have all the components and capabilities of system 10, and can include additional component and capabilities described herein. For example, testing system 310 can include machine-readable storage medium and/or one or multiple hardware/computer/data processors. The term “processor” as used in this disclosure can include one processor or multiple processors for performing executable instructions as described. The performance of the executable instructions can be by different processors distant from one another, by multiple processors working together, or by one processor executing all instructions. Additionally, components of similar name and/or reference number of system 10 and testing system 310 (while adding a “3” to the front of reference numbers of testing system 310) can have similar configurations and/or capabilities. For example, main AI model 48A of system 10 can be similar in configuration and capabilities to test AI model 348A of testing system 310. Similarly, configuration file 14 of system 10 can be similar in configuration, capabilities, etc. as test configuration file 314 of testing system 310.
Testing system 310 utilizes machine learning model 356 in a testing loop to determine test preprocessing parameters in test configuration file 314 that, when applied to raw video data 20, formats test video data 321 such that when test AI model 348A processes 348 test video data 321, test output 350 determined by test AI model 348A is an improvement over main output 50 of main system 10. Test output 350 can include dependent variables that are one or more of the following: correctly identified inferences by test AI model 348A, accuracy of test AI model 348A in identifying inferences, and latency of test AI model 348A. Testing system 310 improves test output 350 as detailed below.
Testing system/pipeline 310 accesses raw video data 20 collected by camera 12 and potentially published at internal topic 42. Raw video data 20 is test preprocessed 344A by testing system 310 to create test video data 321 having scene 324A, which can be a subset of entire frame 22 as shown in
Test preprocessing parameters in test configuration file 314 along with corresponding test output 350 can be received/accessed by machine learning model 356 and/or added to database 358. Database 358 can potentially include other information, such as the time of day the testing loop was performed by testing system 310 (e.g., the preprocessing and processing). The information in database 358 can be used by training module 362 and/or implementation module 360 to determine new test preprocessing parameters of test configuration file 314 that return a greater improvement of test output 350 than the previous test preprocessing parameters of the previous test configuration file 314.
Additionally, testing system 310 can compare test output 350 to the baseline criterion to determine if test output 350 satisfies the baseline criterion. If so, main configuration file 14 can be altered so that preprocessing parameters of main configuration file 14 are similar to test preprocessing parameters of test configuration file 314. Then, main system/pipeline 10 preprocesses 44A raw video data 20 to create video data 21 that is eventually analyzed/processed 48 by main AI model 48A to produce main output 50 that is improved (e.g., increased correct identification of inferences, increased accuracy, and/or decreased latency) as compared to the previous main output 50.
If test output 350 does not satisfy the baseline criterion, machine learning model 356 can alter the test preprocessing parameters in test configuration file 314 such that the new test preprocessing parameters are predictive to produce new test output 350 that satisfies the baseline criterion. The new, altered test preprocessing parameters of test configuration file 314 are then reapplied to new incoming raw video data 20. New test preprocessing 344A is performed to create new test video data 321, which is then analyzed/processed by test AI model 348A to create new test output 350. Testing system 310 then compares new test output 350 to the baseline criterion. If testing system 310 determines that the new test output 350 satisfies the baseline criterion (this can be performed by machine learning model 356), testing system 310 can alter the main preprocessing parameters in main configuration file 14 of system 10 to be similar to the new test preprocessing parameters of the new test configuration file 314. If new test output 350 does not satisfy the baseline criterion, the test preprocessing parameters can be altered and the testing loop can be performed again. This testing loop can continue until the test output satisfies the baseline criterion.
Training module 362 is used to initially identify/select the test preprocessing parameters defined within test configuration file 314 by implementation module 360 (i.e., the test preprocessing parameters that are first applied to raw video data 20 at the beginning of the testing loop). Training module 362 can use information in database 358 to run training sessions similar to the testing loop to refine the accurate prediction of test preprocessing parameters. The information in database 358 can include previously used test preprocessing parameters, set out as independent variables, and the corresponding test outputs 350, set out as dependent variables. With test outputs 350 (as well as the proximity of test outputs 350 to the baseline criterion) and the corresponding test preprocessing parameters known, implementation module 362 can determine which test preprocessing parameters most closely correlate/influence test outputs 350. With enough training datasets in database 358, implementation module 362 can identify the test preprocessing parameters that are most predictive of a change in test output 350 and that would produce a test output 350 that satisfies the baseline criterion and is a greater improvement upon previous test outputs 350 (and main output 50).
Testing system 310 can include machine-readable storage media and/or can include one or multiple hardware/computer/data processors. Testing system 310 can include other components not expressly disclosed herein but that are suitable for performing the functions of testing system 310 and associated methods of optimizing the preprocessing parameters to create video data 21 to improve the analysis/processing by main AI model 48A. For example, testing system 310 can include communication software and/or hardware for pushing/sending optimal preprocessing parameters to main configuration file 14 for use in system 10. One, multiple, or all of the components of testing system 310 can be incorporated into one or multiple hardware/computer/data processors.
Test configuration file 314 is similar in configuration, functionality, and capabilities to configuration file 14 described with regards to
Test configuration file 314 is applied to raw video data 20 (i.e., raw video data 20 is preprocessed 344A) to create test video data 321 having test scene 324A. Test scene 324A can have the same field of view as scene 24A of main system/pipeline 10 so that processing by main AI model 48A and by test AI model 348A can have the same baseline for analysis in creating main output 50 and test output 350, respectively.
After applying test preprocessing parameters defined within test configuration file 314 to raw video data 20 to create test video data 321, test video data 321 can be published to test unified endpoint 318. Test unified endpoint 318 can be the same in configuration, capabilities, and functionality as unified endpoint 18 of system 10, and even could be the same component so as to be one in the same with unified endpoint 18 such that video data 21 and test video data 321 are published at the same location to allow access by main AI model 48A and/or test IA model 348A. In addition to unified endpoint 318, test video data 321 can be published/sent first to a proxy location (such as proxy location 44) and then relayed to unified endpoint 318.
After being published/sent to unified endpoint 318, test video data 321 (with test scene 324A) is available to be subscribed to and accessed by test AI model 348A. Test AI model 348A processes 348 test video data 321 to determine at least one test output 350 indicative of at least one test inference dependent upon test video data 321. Test AI model 348A and main AI model 48 can be similar in configuration, capabilities, and functionality and may even be the same component. Main AI model 48A and test AI model 348A should process video data 21 and test video data 321, respectively, in a similar manner to ensure that the independent variables in determining main output and test output (i.e., the differences between main system/pipeline 10 and testing system/pipeline 310) are only the main preprocessing parameters (in configuration file 14) and the test preprocessing parameters (in test configuration file 314) as opposed to any differences in the way main AI model 48A and test AI model 348A perform the processing. Main AI model 48A and test AI model 348A can be in communication with one another so that if any changes in the processing/analysis are made to main AI model 48A, those changes are also applied to test AI model 348A.
Main AI model 48A and/or test AI model 348A can be a program/model that may have machine learning and can use the video data to determine at least one main output 50 or test output 350 indicative of an inference dependent upon the video data. The inference, for example, can be the amount of a specific product that is viewable in scene 24A or test scene 324A over a defined period of time. Main AI model 48A and/or test AI model 348A can also be, for example, a program/model that determines how many people appear in scene 24A or test scene 324A over a defined period of time. Main AI model 48A and/or test AI model 348A can include other capabilities and/or configurations.
Main output 50 and test output 350 can each include a variety of information dependent upon video data 21 and test video data 321, respectively. That information can include a value representative of a number of correctly identified inferences (e.g., a number of correctly identified products), a value representative of the accuracy of main output 50 or test output 350 (e.g., a number between 0 and 1 showing a probability that the identified inferences are correct), and/or a value representative of the latency of the processing by main AI model 48A and test AI model 348A to identify inferences (e.g., an amount of time the AI model takes to analyze the video data and identify an inference). Main output 50 can be accessed by and/or sent to subscriber 26 for further analysis, use, and/or recordation. Additionally, main output 50 and test output 350 can be accessed by and/or sent to machine learning model 356 for use in determining if and to what extent to alter test preprocessing parameters in test configuration file 314.
Machine learning model 356 can include one or multiple hardware/computer/data processors capable of receiving information regarding test preprocessing parameters, test outputs 350, main outputs 50, baseline criterion, and/or other information and identifying test preprocessing parameters that are predictive to produce improved test outputs 350. Machine learning model 356 can include test configuration file 314, which includes test preprocessing parameters that are applied to raw video data 20 to create test video data 321, and database 358, within which information (e.g., datasets) for identifying test preprocessing parameters is stored and accessed. Additionally, machine learning model 356 can include training module 362, which can include a hardware/computer processor and training software code that determines which test preprocessing parameters are most predictive of a change in test output 350, and implementation module 360, which can be a hardware/computer processor and software code that uses the information determined by training module 362 identify test preprocessing parameters in test configuration file 314 and applies the test preprocessing parameters to perform a testing loop. Moreover, one or multiple components of machine learning model 356 can then be configured to alter the main preprocessing parameters in main configuration file 14 depending on test output 350 and the proximity of test output 350 to the baseline criterion. Thus, machine learning model 356 can be in wired or wireless communication with any components of main system/pipeline 10 and/or testing system/pipeline 310. Machine learning model 356 can be similar in configuration, capabilities, and functionality as gateway 16, and can be the same component as gateway 16 such that any computing/processing/execution by machine learning model 356 can be performed by gateway 16. Machine learning model 356 can include one or multiple storage media for storing information, such as the test preprocessing parameters and/or test configuration file 314, the baseline criterion, main output 50, test output 350, the information in database 358, and/or other information. Machine learning model 356 can be configured to determine and store any algorithms, neural networks, or other predictive systems/methods that determine preprocessing parameters predictive to produce improved test outputs 350. Machine learning model 356 can be located at various locations, including within, adjacent to, or distant from gateway 16, main AI model 48A, test AI model 348A, and/or any other components of main system/pipeline 10 and testing system/pipeline 310.
Testing system/pipeline 310 can include a component similar to gateway 16 to perform all of the functions of gateway 16 with regards to main system/pipeline 10, such as preprocessing 344A raw video data 20 according to test preprocessing parameters in test configuration file 314 to create test video data 321. Alternatively, preprocessing 344A can be performed by machine learning model 356 such that machine learning model 356 can be in communication with internal topic 42 to receive/access raw video data 20 and in communication with test unified endpoint 318 to which test video data 321 is sent/published for access by test AI model 348A. Testing system/pipeline 310 and machine learning model 356 can be the same component/hardware/computer processor.
Machine learning model 356 can receive test output 350 (and main output 50) and compare test output 350 to the baseline criterion. The comparison of test output 350 to the baseline criterion can be a comparison of any data included in test output 350 (such as values representative of correctly identified inferences, accuracy, and/or latency) to the baseline criterion.
The baseline criterion can be established/set through various methods. The baseline criterion can be a value or values designated by a user. For example, a user can select an accuracy value of 0.95 for the baseline criterion such that test output 350 must have an accuracy value at or above 0.95 to satisfy the baseline criterion. In another example, the user can select a latency value to be 10 millisecond for the baseline criterion such that test output 350 must have a latency value at or below 10 millisecond to satisfy the baseline criterion. In another example, the user can select a value of correctly identified inferences to be 12 per minute for the baseline criterion such that test output 350 must have a value of correctly identified inferences to be at or more than 12 per minute to satisfy the baseline criterion. The baseline criterion can include one or more than one of these designated values such that test output 350 must satisfy each of the designated baseline values to satisfy the baseline criterion. Additionally, the baseline criterion can be set by machine learning model 356 depending on a variety of factors, including the values of a previous test output 350.
Moreover, the baseline criterion can be set by the respective values of main output 50. In this situation, test output 350 is, for all intents and purposes, compared to main output 50 because main output 50 establishes the baseline criterion. For example, main output 50 can have an accuracy value of 0.90, which is then set as the baseline criterion. Thus, the baseline criterion is satisfied when a value representative of accuracy of test output 350 is greater than the value representative of the accuracy of main output 50, which is 0.90. In another example, main output 50 can have a latency value of 15 milliseconds (i.e., a time it takes main AI model 48A to process video data 21 to determine main output 50), which is then set as the baseline criterion. Thus, the baseline criterion is satisfied when a value representative of latency of test output 350 (i.e., a time it takes test AI model 348A to process test video data 321 to determine test output 350) is less than the value representative of latency of main output 50, which is 15 milliseconds. In another example, main output 50 can have a value of correctly identified inferences of 10 per minute, which is then set as the baseline criterion. Thus, the baseline criterion is satisfied when a value representative of correctly identified inferences by test AI model 348A (as set out in test output 350) is greater than the value representative of correctly identified inferences by main AI model 348A (as set out in main output 50), which is 10 per minute. The baseline criterion can include one or more than one of these values based on main output 50 such that test output 350 must satisfy each of the designated baseline values to satisfy the baseline criterion.
As mentioned above, machine learning model 356 compares test output 350 to the baseline criterion. If test output 350 satisfies the baseline criterion, machine learning model 356 can be configured to alter the preprocessing parameters in main configuration file 14 to be the same as the test preprocessing parameters in test configuration file 314 so that main output 50 will be similar to test output 350 having the improved output (after preprocessing and processing of video data by main system/pipeline 10). Thus, main output 50 is improved/optimized through machine learning model 356 and the testing performed by testing system/pipeline 310. After the preprocessing parameters defined in main configuration file 14 are updated by testing system/pipeline 310, testing system/pipeline 310 can continue the testing loop by implementation module 360 and/or training module 362 to find even more improved/optimal preprocessing parameters. Information from each testing loop (e.g., test preprocessing parameters, test outputs 350, and time-of-day information) can be added to database 358 for further training of machine learning model 356.
If test output 350 fails to satisfy the baseline criterion, machine learning model 356 can alter the test preprocessing parameters in test configuration file 314. Then, the testing loop can be repeated by preprocessing 344A raw video data 20 according to the newly altered test preprocessing parameters in newly altered test configuration file 314, creating new test video data 321, publishing/sending new test video data 321 to test unified endpoint 318, and accessing and processing 348A by test AI model 348A to create new test output 350. Machine learning model 356 can then compare new test output 350 to the same or a different baseline criterion. This testing loop can be repeated as many times as desired by machine learning model 356 to determine test preprocessing parameters that are predictive of an improved test output 350, such as until test output 350 satisfies the baseline criterion or even after test output 350 has satisfied the baseline criterion and the main preprocessing parameters have been altered to be similar to the most current test preprocessing parameters.
Similar to main preprocessing parameters of main configuration file 14, test preprocessing parameters of test configuration file 314 can include multiple test video edits (see the discussion regarding
Main system/pipeline 10 and testing system/pipeline 310 operate simultaneously such that main system 10 analyzes raw video data 20 by receiving raw video data 20 from camera 12, preprocessing 44A raw video data 20 according to the preprocessing parameters defined within configuration file 14 with preprocessing 44A including formatting raw video data 20 to create video data 21, accessing video data 21 by main AI model 48A, and processing 48 video data 21 by main AI model 48A to determine main output 50 that is indicative of an inference dependent upon video data 21. Testing system/pipeline 310 improves main output 50 via machine learning model 356 by receiving raw video data 20 from camera 12, identifying test preprocessing parameters in test configuration file 314 based upon database 358 with the test preprocessing parameters being predictive to produce test output 350 that satisfies the baseline criterion, preprocessing 344A raw video data 20 according to the test preprocessing parameters with the test preprocessing 344A including formatting raw video data 20 to create test video data 321, accessing test video data 321 by test AI model 348A, and processing 348 test video data 321 by test AI model 348A to determine test output 350 that is indicative of a test inference dependent upon test video data 321. Testing system/pipeline 310 can compare test output 350 to the baseline criterion and, in response to test output 350 satisfying the baseline criterion, testing system 310 can alter main preprocessing parameters in configuration file 14 to be similar to test preprocessing parameters in test configuration file 314. In response to test output 350 failing to satisfy the baseline criterion, machine learning model 356 identifies new test preprocessing parameters based upon the first test preprocessing parameters and database 358 with the new test preprocessing parameters within test configuration file 314 being predictive to produce a new test output 350 that satisfies the baseline criterion. Testing system/pipeline 310 repeats the steps of preprocessing, processing, and comparing the new test output 350 to the baseline criterion. After each testing loop, each of the test preprocessing parameters and the corresponding test output 350 are added to database 358 to provide an additional dataset used in training machine learning model 356.
For training of machine learning model 356 to improve main output 50, machine learning model 356 can include a hardware/computer processor and machine learning training software code stored in memory with the hardware/computer processor configured to execute the machine learning training software code. Machine learning model 356, which includes training module 362, receives raw video data 20 from camera 12 (or uses previously stored raw video data that is stored for the purpose of training machine learning model 356). Machine learning model 356 identifies, by the machine learning training software code as executed by the hardware/computer processor, first test preprocessing parameters having video edits 38 defined within test configuration file 314 with the first test preprocessing parameters having independent variables that are at least one of the following: the video edits, a value of each of the video edits, and an order in which the video edits are performed on raw video data 20. Training machine learning model 356 then includes preprocessing 344A raw video data 20 according to the first test preprocessing parameters to create first test video data 321, accessing first test video data 321 by test AI model 348A, and processing 348 first test video data 321 by test AI model 348A to determine first test output 350 that is indicative of a first inference dependent upon first test video data 321 with first test output 350 having dependent variables that are at least one of the following: correctly identified inferences by test AI model 348A, accuracy of test AI model 348A, and latency in determining the test inference by test AI model 348A. The first test preprocessing parameters and first test output 350 are added to database 358 to provide an additional dataset used in training machine learning model 356. Then, second test preprocessing parameters are identified based on the first test preprocessing parameters and first test output 350 with the second test preprocessing parameters having different values of independent variables than the first test preprocessing parameters, raw video data 20 is preprocessed 344A according to the second test preprocessing parameters within test configuration file 314 to create second test video data 321, and second test video data 321 is processed 348 by test AI model 348A to determine second test output 350 that is indicative of a second inference dependent upon second test video data 321 with second test video data 321 having dependent variables. The second test preprocessing parameters and second test output 350 are added to database 358 to provide an additional dataset used in training machine learning model 356. Training module 362 can analyze the information in database 258 and determine which independent variables (e.g., which video edits of the preprocessing parameters) are most influential of the dependent variables (e.g., the outputs). Based on information in database 358, including the independent variables from each of the test preprocessing parameters and the dependent variables from each test output 350, machine learning model 356 extracts main preprocessing parameters having independent variables that are predictive of a change in the dependent variables.
Training machine learning model 356 can further include continually adjusting the testing preprocessing parameters to create further test video data 321, processing 348 the further test video data 321 by test AI model 348A to determine further test outputs 350 indicative of further inferences dependent upon the first test video data 321 with the first test outputs 350 having dependent variables, and adding the test preprocessing parameters and corresponding further test outputs 350 to database 358. Once machine learning model 356 (e.g., training module 362) has been trained to an acceptable level such that preprocessing parameters have been identified that are predictive to produce main output 50 that satisfies the baseline criterion (and/or is an improvement upon the previous main output 50), machine learning model 356 (e.g., implementation module 360) is used in conjunction with main system/pipeline 10 by altering the main preprocessing parameters in main configuration file 14 to be the improved preprocessing parameters. Then, main system/pipeline 10 receives raw video data 20, preprocesses 44A raw video data 20 according to the altered main preprocessing parameters in main configuration file 14 to create main video data 21, and processes 48 main video data 21 by main AI model 48A to determine main output 50 that is indicative of a main inference dependent upon main video data 21 with main output 50 potentially satisfying the baseline criterion and/or being an improvement upon previous main outputs 50. While main system/pipeline 10 is preprocessing and processing video data, machine learning model 356 (e.g., implementation module 360) can perform testing loops on raw video data 20 to collect more datasets for database 358 and improve the prediction of preprocessing parameters.
While the invention has been described with reference to an exemplary embodiment(s), it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted for elements thereof without departing from the scope of the invention. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the invention without departing from the essential scope thereof. Therefore, it is intended that the invention not be limited to the particular embodiment(s) disclosed, but that the invention will include all embodiments falling within the scope of the appended claims.
