Patent Application
20210097290
This patent document relates generally to systems and methods for retrieving video. Examples of retrieving video in feature descriptor domain in an artificial intelligence semiconductor solution are provided.
In video analysis and other applications, such as video retrieval, processing of image pixels is often performed. This requires high computing power because of large amount of information in image pixels. For example, a one-hour video captured in 30 frames per second may contain 108,000 image frames. If the video resolution is the standard VGA at 640×480, the number of pixels in the video w ill amount to 30 billion pixels. Some existing systems extract key frames from a video before performing further analysis, so that the computation is limited to processing key frames instead of all of the image frames in the video. Key frame detection generally determines the image frames in a video where an event has occurred. The examples of an event may include a motion, a scene change or other condition changes in the video. Key frame detection generally processes multiple image frames in the video and may still require extensive computing resources. Other technologies may include selecting a subset of image frames in a video either at a fixed time interval or a random time interval, without assessing die content of the images in the video. However, these methods may be less than ideal because the image frames selected may not be the true key frames that reflect when an event occurs. In other words, a randomly selected key frame may be redundant to a previous key frame, thus the randomly selected key frame does not provide any valuable information. Further, whether key frame based or non-key frame based, video retrieval may require comparing image frames (e.g., in a query video) to image frames (e.g., in a video database). This comparing process is based on processing image pixels and thus requires large computations.
This document is directed to systems and methods for addressing the above issues and/or other issues.
The present solution will be described with reference to the following figures, in which like numerals represent like items throughout the figures.
FIG. A illustrates a flow diagram of an example process of retrieving video from a video database in accordance with various examples described herein.
As used in this document, the singular forms “a”, “an”, and “the” include plural references unless the context clearly dictates otherwise. Unless defined otherwise, all technical and scientific terms used herein have the same meanings as commonly understood by one of ordinary skill in the art. As used in this document, the term “comprising” means “including, but not limited to.”
Examples of “artificial intelligence logic circuit” or “AI logic circuit” include a logic circuit that is configured to execute certain AI functions such as a neural network in AI or machine learning tasks. An AI logic circuit can be a processor. An AI logic circuit can also be a logic circuit that is controlled by an external processor and executes certain AI functions.
Examples of “integrated circuit,” “semiconductor chip.” “chip,” or “semiconductor device” include an integrated circuit (IC) that contains electronic circuits on semiconductor materials, such as silicon, for performing certain functions. For example, an integrated circuit can be a microprocessor, a memory, a programmable array logic (PAL) device, an application-specific integrated circuit (ASIC), or others. An integrated circuit that contains an AI logic circuit is referred to as an AI integrated circuit.
Examples of an “AI chip” include hardware- or software-based device that is capable of performing functions of an AI logic circuit. An AI chip can be physical or virtual. For example, a physical AI chip may include an embedded cellular neural network, which may contain weights and/or parameters of a convolution neural network (CNN) model. A virtual AI chip may be software-based. For example, a virtual AI chip may include one or more processor simulators to implement functions of a desired AI logic circuit of a physical AI chip.
Examples of “AI model” include data that include one or more weights that, when loaded inside an AI chip, are used for executing the AI chip. For example, an AI model for a given CNN may include the weights, biases, and other parameters for one or more convolutional layers of the CNN. Here, the weights and parameters of an AI model are interchangeable.
In some examples, the system 100 may also include a feature extractor 112 configured to extract one or more feature descriptors from multiple images in a candidate video in a video database. Similar to the feature extractor 104, examples of a feature descriptor from the feature extractor 112 may also include any values that are representative of one or more features of an image. For example, the feature descriptor may include a vector containing values representing multiple channels of a feature map of any of the multiple images in the candidate video. In a non-limiting example, an input image of the CNN may have 3 channels, whereas the feature map from the CNN may have 512 channels. In such case, the feature descriptor may be a vector having 512 values. The output feature descriptors from the feature extractor 112 may include multiple vectors from the multiple images in a video in the video database. In a non-limiting example, a candidate video in the video database may be fed to the feature extractor 112 to generate the feature descriptors for the candidate video. In some examples, the feature extractors 104, 112 may be implemented in a CNN, which will be further described in the present disclosure.
With further reference to
In some examples, the system 100 may access multiple image frames, e.g., a sequence of image frames, of the query video or the candidate video in the video database. For example, the system may access the query video or the candidate video stored in a memory or on the cloud over a communication network (e.g., the Internet), and extract the sequence of image frames in the video. In some or other scenarios, the system may receive a query video or a plurality of image frames directly from an image sensor. The image sensor may be configured to capture a video or an image. For example, the image sensor may be installed in a video surveillance system and configured to capture video/images of a vehicle exiting a garage, a parking lot, or a building. The system 100 may be configured to search the previous stored surveillance video in a video database to retrieve a similar video of the same vehicle to verify that the video exiting the garage had previously entered the same garage.
Optionally, the system 100 may further include compression systems 102, 110, configured to respectively reduce the sizes of the plurality of image frames in the query video and the candidate video to a proper size so that the plurality of image frames are suitable for uploading to a CNN model for implementing the feature extractor. In some examples, the CNN model may be executed in a physical A1 chip having hardware constraints. For example, the A1 chip may include a buffer for holding input images up to 224×224 pixels for each channel. In such case, the compression systems 102, 110 may reduce each of the image frames to a size at or smaller than 224×224 pixels. In a non-limiting example, the compression systems 102, 110 may down sample each image frame to the size constrained by the AI chip. Additionally, and/or alternatively, the compression systems 102. 110 may crop each of the plurality of image frames in a video to generate multiple instances of cropped images. For example, for an image frame having a size of 640×480, the instances of cropped images may include one or more sub-images, each of the sub-images being smaller than the original image and cropped from a region of the original image. In a non-limiting example, the system may crop the input image in a defined pattern to obtain multiple overlapping sub-images which cover the entire original image. In other words, each of the cropped images may contain image contents attributable to a feature descriptor based on each cropped image. Accordingly, for an image frame, the feature extractor 104, 112 may access multiple instances of cropped images and produce a feature descriptor based on the multiple instances of cropped images. The details will be further described with reference to
With further reference to
In some examples, the invariant pooling layer 208 may be configured to determine a feature descriptor based on the feature maps obtained from the CNN. The pooling layer 208 may include a square-root pooling, an average pooling, a max pooling or a combination thereof. The CNN may also be configured to perform a region of interest (ROI) sampling on the feature maps to generate multiple updated feature maps. The various pooling layers may be configured to generate a feature descriptor for various rotated images.
Additionally, each of the feature maps from various image rotations may be nested to include multiple cropped images (regions) from the input image. The cropped images may be fed to the CNN to generate multiple feature maps, each of the feature maps representing a cropped region. The feature extractor may further concatenate (stack) the features maps from multiple cropped images nested in each set of feature maps from air image rotation. In other words, each feature map from a rotated image may include a set of feature maps comprising multiple feature maps that are concatenated (stacked together), where each feature map in the set results from a respective cropped image from a respective rotated image. As the cropped images from an input image (or rotated input image) may have different sizes, the feature maps within each set of feature maps may also have different sizes.
Additionally, and/or alternatively, a region of interest (ROl) sampling may be performed on top of each set (stack) of feature maps. Various ROl methods may be used to select one or more regions of interest from each of the feature maps. Thus, a feature map in the set of feature maps for an image rotation may be further nested to include multiple sub-feature maps, each representing a ROl within that feature map. For example, an image of a size of640x480 may result in a feature map of a size of 20×15. In a non-limiting example, the feature extractor 300 may generate two ROl samplings, each having a size of 15×15, where the two ROl samplings may be overlapping, covering the entire feature map. In another non-limiting example, the feature extractor 300 may generate six ROl samplings, each having a size of 10×10, where the six ROl samplings may be overlapping to cover the entire feature map. All of die feature maps for all image rotations and the nested sub-feature maps for ROls within each feature map may be concatenated (stacked together) tor performing the invariance pooling.
In some examples, the invariance pooling 314 may be a nested invariance pooling and may include one or more pooling operations. For example, the invariance pooling 314 may include a square-root pooling 316 performed on the ROIs of all concatenated feature/sub-feature maps to generate a plurality of values 308, each representing the square-root values of the pixels in the respective ROl. Further, the invariance pooling 314 may include an average pooling 318 to generate a feature vector 310 for each set of feature maps (corresponding to each image rotation, e.g., at 0, 90, 180 and 270 degrees, respectively), each feature vector corresponding to an image rotation and based on an average of the square-root values from multiple sub-feature maps. Further, the invariance pooling 314 may include a Max pooling 320 to generate a single feature descriptor 312 based on the maximum values of the feature vectors 310 obtained from the average pooling. As shown, for each of a plurality of image frames of a video segment, the feature extractor may generate a corresponding feature descriptor, such as 312. In a non-limiting example, the feature descriptor may include a one-dimensional (1D) vector containing multiple values. The number of values in the 1D descriptor vector may correspond to the number of output channels in the CNN.
In some examples, the process 400 may not need to compare every single image frame in the candidate video with the query video. Instead, the process 100 may set a reference frame in the candidate video and determine a similarity between the reference frame and the query video in the manner as described above. The process 400 may subsequently compare each succeeding image frame in the candidate video with the reference frame. If the succeeding image frame is similar to the reference frame, the process 400 may determine the similarity between the succeeding image frame and the query video based on the similarity between the reference frame and the query video, instead of computing the similarity between the succeeding image frame and the query video as described above. If the succeeding image frame and the reference frame are not similar, the process 400 may reset the succeeding image frame to the reference frame and compute the similarity between the reference frame and the query video in the above described manner.
Now with further reference to
In determining the distance between the reference frame and a respective image frame in the query video, in some example, the process may determine a distance value between the feature descriptor of the reference frame in the candidate video and the feature descriptor of the respective image frame in die query video, both of the feature descriptors are provided from the feature extractors 112 and 104, respectively. In anon-limiting example, the feature descriptor may be a ID vector. For example, if the output of the CNN implementing the feature extractor (e.g., 202 in
In a non-limiting example, the distance between a first feature descriptor u and a second feature descriptor v may be expressed as:
where u⋅v is the dot product of u and v and ∥u∥2 and ∥v∥a are Euclidean norms. In an example, if u and v have the same direction, then the cosine distance may have a minimal value, such as zero. If u and v are perpendicular to each other, then the cosine distance may have a maximum value, e.g., a value of one. In here, the distance value between two feature descriptors corresponding to two image frames may indicate the extent of changes between the two image frames. A higher distance value may indicate a more significant difference between the two corresponding image frames (which may indicate an occurrence of an event) than a lower distance value docs.
With further reference to box 401, in initializing the current frame, the process 400 may select the current frame as the next image frame succeeding the reference frame In some examples, the current frame may skip a number of image frames (e.g., an integer n) after the reference frame. For example, the current frame may be the reference frame+n. In other words, the process 400 may skip n image frames assuming that the image frames within the neighborhood of n frames are similar and do not need to be processed. The integer n may be any suitable number. For example, n may have a value of 10, 15,16, or other values.
With further reference to
Returning to box 406, if the distance between the current frame and the reference frame equals or exceeds the threshold T1, it means that the current frame is likely not close to the reference frame and cannot be represented by the reference frame. Then, the process 400 may compare the current frame with the query video to determine the distances between the current frame and respective image frames in the query video at 408 in a similar manner as described in determining the similarity between the reference frame and the query video in 401. In some examples, if the average of the distances between the current frame and the respective image frames in the query video is below a threshold, e.g., T2, at 410, then the process 400 may determine that the current frame is similar to the query video at 412. Otherwise, the process 400 may determine that the current frame is not similar to the query video at 414. At this point, the distance between the current frame and the query video has just been calculated instead of inherited from the distance between the reference frame and the query video. In various embodiments, the process 400 may also use other combinations of distances to determine the similarity between the current frame and the query video. For example, the process 400 may determine whether a maximum distance or a median distance is below the threshold T2 at 410 and proceed to boxes 412 or 414 depending on the determination.
With continued reference to
With further reference to
Returning to
With further reference to
In a non-limiting example, determining the distance values between two sets of feature descriptors may include calculating a distance value between a feature descriptor pair containing a feature descriptor from the first set and a corresponding feature descriptor from the second set. In the example above, the first set of feature descriptors may include 10 vectors each corresponding to an image frame between I-10 and the second set of feature descriptors may include 10 vectors each corresponding to a respective image frame between 11-20. Then, the process of determining the distance values between the first and second sets of feature descriptors may include determining multiple distance values. For example, the process may determine a first distance value between the feature descriptor corresponding to image frame 1 (from the first set) and the feature descriptor corresponding to image frame 11 (from the second set). The process my determine the second distance value based on the descriptor corresponding to image frame 2 and the descriptor corresponding to image frame 12. The process may determine other distance values in a similar manner.
In some examples, in determining the distance value, the process 506 may use a cosine distance. For example, if a vector in the first set of feature descriptors is w, and the corresponding vector in the second set of feature descriptors is v, then the cosine distance between vectors u and
where u⋅v is the dot product of u and v and |u|2 and |v|2 are Euclidean norms. In an example, if u and v have the same direction, then the cosine distance may have a minimal value, such as zero. If u and v are perpendicular to each other, then the cosine distance may have a maximum value, e.g., a value of one. In here, the distance value between two feature descriptors corresponding to two image frames may indicate the extent of changes between the two image frames. A higher distance value may indicate a more significant difference between the two corresponding image frames (which may indicate an occurrence of an event) than a lower distance value does. In other words, if a distance value between two feature descriptors exceeds a threshold, the system may determine that an event has occurred between the corresponding image frames. For example, the event may include a motion ill the image frame (e.g., a car passing by in a surveillance video) or a scene change (e.g., a camera installed on a vehicle capturing a scene change when driving down the road), or change of other conditions. In such case, the process may determine that the image frames where the significant changes have occurred in the corresponding feature descriptors be key frames. Conversely, a lower distance value between the feature descriptors of two image frames may indicate less significant change or no change between the two image frames, which may indicate that the two image frames contain static background of the image scenes. In such case, the process may determine that such image frames are not key frames.
With further reference to
In a non-limiting example, the process 514 may select the key frames from the top feature descriptors which resulted in distance values exceeding the threshold. In the example above, if the feature descriptors of image frames 14 and 15 are above the threshold, then the process 514 may determine that image frames 14 and 15 are key frames. Additionally, and/or alternatively, if the feature descriptors of multiple image frames in the second subset of image frames have exceeded the threshold, the process may select one or more top key frames whose corresponding feature descriptors have yielded highest distance values. For example, between image frames 14 and 15, the process may select frame 15, which yields a higher distance value than frame 14 docs In another non-limiting example, if image frames 11, 12, 14, 15 all yield distance values above the threshold, the process may select all of these image frames as key frames. Alternatively, the process may select two key frames whose feature descriptors yield the two highest distance values.
Now the first and second sets of feature descriptors are processed, the process 500 may move to process additional feature descriptors. In some examples, the process 500 may update a feature descriptor access policy at 510. 516 depending whether one or more key frames are detected. For example, if one or more key frames are detected at 514, the process 516 may update the first set of feature descriptors to include the current second set of feature descriptors, and update the second set of feature descriptors to include additional feature descriptors corresponding to a third subset of image frames of the plurality of image frames. In the above example, the first set of feature descriptors may be updated to include the second set of feature descriptors, such as the feature descriptors corresponding to image frames 11-20; and the second set of feature descriptors may be updated to include a new set of feature descriptors corresponding to image frames 21-30. In such case, subsequent distance values between tire first and second sets of feature descriptors may be determined based on the feature descriptors corresponding to image frames 11-20 and 21-30, respectively.
Alternatively, if no key frames are detected at 514, then the process 510 may update the second set of feature descriptors to include additional feature descriptors corresponding to a third subset of image frames of the plurality of image frames. For example, if no key frames are detected in image frames 11 -20, then the second set of feature descriptors may include feature descriptors corresponding to the new set of image frames 21-30. In some examples, the First set of feature descriptors may remain unchanged. For example, the First set of feature descriptors may remain the same and correspond to image frames 1-10. Alternatively, the First set of feature descriptors may be set to one of the feature descriptors. For example, the First set of feature descriptors may include (lie feature descriptor corresponding to image frame 10, In such case, subsequent distance values between the First and second sets of feature descriptors may be determined based on the feature descriptor corresponding to image frame 10 and feature descriptors corresponding to image frames 21-30. In other words, the image frames 11-20 are ignored.
In some examples, the process 500 may repeat blocks 506-516 until the process determines that the feature descriptors corresponding to all of the plurality of images frames in the video segment have been accessed at 518. When such determination is made, the process 500 may store the key frames at 520. Otherwise, die process 500 may continue repeating 506-516. in some variations, block 520 may be implemented when all feature descriptors have been accessed at 518. Alternatively, and/or additionally, block 520 may be implemented as key frames are detected (e.g., at 514) in one or more of the iterations. As described above with respect to
Returning to
In a non-limiting example, the first threshold may be 100 image frames, the second threshold may be 10,000 image frames, and the value n may be 20. In this case, if the number of image frames in a video segment is less than 100, the system may process the entire video without detecting key frames. If the number of image frames in the video segment is between 100 and 10,000, the system may detect key frames in the video segment to determine at least 100 key frames. Alternatively, if the number of image frames in the video segment exceeds 10,000, the system may apply a more aggressive key frame detection so that the remaining key frames from key frame detection is about 10.000/20=500. It is appreciated that the first threshold value, the second threshold value, and/or the variable n may vary.
As described with respect to
It will be readily understood that the components of the present solution as generally described herein and illustrated in the appended figures in
An optional display interface 630 may permit information from the bus 600 to be displayed on a display device 635 in visual, graphic, or alphanumeric format. An audio interface and audio output (such as a speaker) also may be provided. Communication with external devices may occur using various communication ports 640 such as a transmitter and/or receiver, antenna, an RFID tag and/or short-range, or near-field communication circuitry. A communication port 640 may be attached to a communications network, such as the Internet, a local area network, or a cellular telephone data network.
The hardware may also include a user interlace sensor 645 that allows for receipt of data from input devices 650 such as a keyboard, a mouse, a joystick, a touchscreen, a remote control, a pointing device, a video input device, and/or an audio input device, such as a microphone. Digital image frames may also be received from an image capturing device 655 such as a video or camera that can either be built-in or external to the system. Other environmental sensors 660, such as a GPS system and or a temperature sensor, may be installed on system and communicatively accessible by the processor 605, either directly or via the communication ports 640. The communication ports 640 may also communicate with the AI chip to upload or retrieve data to/from the chip. For example, a processing device on the network may be configured to perform operations in the image sizing unit (FIG. I) to and upload the image frames to the AI chip for performing feature extraction via the communication port 640. Optionally, the processing device may use an SDK (software development kit) to communicate with the AI chip via the communication port 640. The processing device may also retrieve the feature descriptors at the output of the AI chip via the communication port 640. The communication port 640 may also communicate with any other interface circuit or device that is designed for communicating with an integrated circuit.
Optionally, the hardware may not need to include a memory, but instead programming instructions are run on one or more virtual machines or one or more containers on a cloud. For example, the various methods illustrated above may be implemented by a server on a cloud that includes multiple virtual machines, each virtual machine having an operating system, a virtual disk, virtual network and applications, and the programming instructions for implementing various functions in the robotic system may be stored on one or more of those virtual machines on the cloud.
Various embodiments described above may be implemented and adapted to various applications. For example, the AI chip having a cellular neural network architecture may be residing in an electronic mobile device. The electronic mobile device may use a built-in AI chip to generate the feature descriptor. In some scenarios, the mobile device may also use the feature descriptor to implement a video search application such as described with reference to
The various systems and methods disclosed in this patent document provide advantages over the prior art, whether implemented, standalone, or combined. For example, by using the feature descriptors to retrieval video, the amount of information for video retrieval are reduced from a two-dimensional array of pixels to 1D vectors. This is advantageous in that the processing associated with video retrieval is done at feature vector level instead of pixel level, allowing the process to take into consideration a richer set of image features while reducing the memory space required for search video at pixel level and computing time. Further, the comparator (e.g., 106 in
Further, the configuration of the feature extractor (e.g., 104,112 in
It will be recognized by those skilled in the art that changes, modifications, or combinations may be made to the above-described embodiments without departing from the broad inventive concepts of the invention. It should therefore be understood that the present solution is not limited to the particular embodiments described herein, but is intended to include all changes, modifications, and all combinations of various embodiments that are within the scope and spirit of the invention as defined in the claims.
