Trailer angle detection using end-to-end learning

Abstract

A method for training an image-based trailer identification system comprises capturing a plurality of captured images in a field of view and identifying a detected trailer angle for a trailer in connection with a vehicle in each of the captured images. The method further comprises comparing the captured images and the corresponding trailer angles to a predetermined image set comprising a plurality of teaching trailer angles and identifying at least one required trailer angle of the teaching trailer angles that is not included in the captured images. Based on the captured images, a simulated angle image is generated. The simulated image comprises a depiction of the trailer in connection with the vehicle at the at least one required angle not included in the captured images. The method further comprises supplying the simulated angle image to the identification system for training.

Description

FIELD OF THE INVENTION

The present invention generally relates to trailer backup assist systems, and, more particularly, to trailer backup assist systems employing trailer angle detection through image processing.

BACKGROUND OF THE INVENTION

Reversing a vehicle while towing a trailer can be challenging for many drivers, particularly for drivers that drive with a trailer on an infrequent basis or with various types of trailers. Some systems used to assist a driver in backing a trailer rely on trailer angle measurements to determine the position of the trailer relative to the vehicle. Thus, the accuracy and reliability of the trailer angle measurements can be critical to the operation of the trailer backup assist system.

SUMMARY OF THE INVENTION

According to one aspect of the present invention, a method for training an image-based trailer identification system is disclosed. The method comprises capturing a plurality of captured images in a field of view and identifying a detected trailer angle for a trailer in connection with a vehicle in each of the captured images. The method further comprises comparing the captured images and the corresponding trailer angles to a predetermined image set comprising a plurality of teaching trailer angles and identifying at least one required trailer angle of the teaching trailer angles that is not included in the captured images. Based on the captured images, a simulated angle image is generated. The simulated image comprises a depiction of the trailer in connection with the vehicle at the at least one required angle not included in the captured images. The method further comprises supplying the simulated angle image to the identification system for training.

The method may further comprise one or more of the following steps alone or in various combinations. The additional steps of the method may include the following:

• • the identifying of the detected trailer angle is determined by digitally processing the image data via a feature extraction technique;
  • the feature extraction technique comprises a least one of a line detection;
  • the identification system comprises a neural network;
  • the plurality of teaching trailer angles comprises a range of trailer angles for identification;
  • the plurality of teaching trailer angles comprises an increment between each trailer angle in the range of trailer angles;
  • detecting a lighting condition in the captured images and comparing the lighting condition to a required lighting condition for training the identification system;
  • in response to the lighting condition of the captured images differing from the required lighting condition, adjusting the lighting of at least one of the captured images to generate the simulated lighting image;
  • supplying the simulated lighting image to the identification system for training;
  • the image data is captured by a plurality of vehicles, each configured to capture the images in the field of view; and/or
  • storing the captured images from the plurality of vehicles and compiling the captured images for comparing the captured images.

According to another aspect of the present invention, a system for training a neural network is disclosed. The system comprises at least one imaging device configured to capture a plurality of captured images of a trailer in connection with a vehicle and an image processing apparatus. The image processing apparatus is configured to identify a detected trailer angle from each of the captured images based on an image processing routine and compare the captured images and the corresponding detected trailer angles to a predetermined image set comprising a plurality of teaching trailer angles. The image processing apparatus is further configured to identify at least one required trailer angle of the teaching trailer angles that is not included in the captured images and generate a simulated angle image based on the captured images. The simulated angle image comprises a depiction of the trailer in connection with the vehicle at the at least one required angle not included in the captured images. The simulated angle image is supplied to the neural network for training.

The system may further comprise one or more of the following elements alone or in various combinations. The additional elements may include the following:

• • the simulated angle image comprises a depiction of the trailer in connection with the vehicle at the at least one required angle not included in the captured images;
  • the image processing routine comprises a digital image processing technique comprising a feature extraction technique;
  • the feature extraction technique comprises a least one of a line detection;
  • the plurality of teaching trailer angles comprises a range of trailer angles for identification;
  • the plurality of teaching trailer angles comprises an increment between each trailer angle in the range of trailer angles;
  • detect a lighting condition in the captured images and compare the lighting condition to a required lighting condition for training the identification system; and/or
  • in response to the lighting condition of the captured images differing from the required lighting condition, adjust the lighting of at least one of the captured images and generate the simulated lighting image.

According to yet another aspect of the present invention, a system for training a neural network is disclosed. The system comprises at least one imaging device configured to capture a plurality of captured images of a trailer in connection with a coupler of a vehicle and an image processing apparatus. The image processing apparatus is configured to identify a first trailer type in the captured images based on a feature extraction technique configured to identify at least one feature of the trailer relative to the coupler and compare the first trailer type to a plurality of teaching trailer types. The image processing apparatus is further configured to generate a simulated trailer image comprising a second trailer type of the plurality of teaching trailer types that is not included in the captured images and supply the simulated trailer image to the neural network for training.

These and other features, advantages, and objects of the present invention will be further understood and appreciated by those skilled in the art by reference to the following specification, claims, and appended drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

In the drawings:

FIG. 1 is a top perspective view of a vehicle attached to a trailer with one embodiment of a trailer angle sensor for operating a trailer backup assist system;

FIG. 2 is a block diagram illustrating one embodiment of the trailer backup assist system;

FIG. 3 is a kinematic model of the vehicle and trailer shown in FIG. 1;

FIG. 4A is a process diagram of a training process for a neural network configured to estimate an angle between a vehicle and a trailer;

FIG. 4B is a process diagram of an operating process for a neural network configured to estimate an angle between a vehicle and a trailer;

FIG. 5A demonstrates examples of cropped images configured as an input to a neural network demonstrating a plurality of trailer types;

FIG. 5B demonstrates examples of cropped images configured as an input to a neural network demonstrating a plurality of lighting or environmental conditions;

FIG. 5C demonstrates examples of cropped images configured as an input to a neural network demonstrating a plurality of trailer angles;

FIG. 6 is a flow chart of a training procedure for generating image data for training a neural network;

FIG. 7A is a projected view of a trailer in connection with a vehicle at a first angle;

FIG. 7B is a projected view of a trailer in connection with a vehicle at a second angle;

FIG. 7C is a projected view of a simulated or augmented trailer angle;

FIG. 8A is a projected view of a trailer in connection with a vehicle demonstrating a first lighting or environmental condition; and

FIG. 8B is a projected view of a simulated or augmented image of the trailer demonstrating a second lighting or environmental condition in accordance with the disclosure.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

For purposes of description herein, it is to be understood that the disclosed trailer backup assist system and the related methods may assume various alternative embodiments and orientations, except where expressly specified to the contrary. It is also to be understood that the specific devices and processes illustrated in the attached drawings, and described in the following specification, are simply exemplary embodiments of the inventive concepts defined in the appended claims. While various aspects of the trailer backup assist system and the related methods are described with reference to a particular illustrative embodiment, the disclosed invention is not limited to such embodiments, and additional modifications, applications, and embodiments may be implemented without departing from the disclosed invention. Hence, specific dimensions and other physical characteristics relating to the embodiments disclosed herein are not to be considered as limiting, unless the claims expressly state otherwise.

As used herein, the term “and/or,” when used in a list of two or more items, means that any one of the listed items can be employed by itself, or any combination of two or more of the listed items can be employed. For example, if a composition is described as containing components A, B, and/or C, the composition can contain A alone; B alone; C alone; A and B in combination; A and C in combination; B and C in combination; or A, B, and C in combination.

Referring to FIGS. 1, 2, and 3, reference numeral 10 generally designates a trailer backup assist system for controlling a backing path of a trailer 10 attached to a vehicle 12. The system may allow a driver of the vehicle 12 to specify a desired curvature of the backing path of the trailer 10. In order to achieve such operation, a trailer angle γ (shown in FIG. 3) between the vehicle 12 and the trailer 10 may be monitored to provide feedback to the system 8 throughout operation. However, accurately detecting the trailer angle γ may be challenging when considering the wide variations in trailer hitch types, weather-related visibility conditions, lighting conditions, trailer angle ranges, and various additional variables that may cause variations in measurement. In order to improve the reliability of identifying the trailer angle γ, the disclosure provides for an improved system and method for end-to-end learning to identify a trailer angle γ.

In particular, the disclosure provides for the detection of the trailer angle γ based on the image data captured by an imaging device 14. Based on the image data captured by the imaging device 14, the system 8 may identify various characteristics of the trailer 10 based on a variety of image processing techniques (e.g. edge detection, background subtraction, template matching etc.). However, due to variations related in the trailer 10 and the local environment (e.g. shadows, textured surfaces, noise, etc.), conventional image processing techniques may not be sufficiently robust to reliably and accurately monitor the trailer angle γ without the aid of additional sensors.

As discussed herein, the improved systems and methods may utilize neural networks to improve the reliability and accuracy of the identified trailer angle γ to improve operation of the system 8. The neural networks and related methods may be configured to learn how to accurately detect the trailer angle γ without human intervention such that the resulting neural network may accurately identify the trailer angle solely based on image data acquired from the imaging device 14. Accordingly, the methods and systems discussed herein, may detect the trailer angle γ reliably without the aid of additional sensors, patterned markers or visual cues, or other aids that may otherwise be required to enable accurate operation of the system 8.

As further discussed in reference to FIGS. 3-6, the invention disclosure provides a solution for the detection of the trailer angle γ based on deep learning and convolutional networks. In this way, the system 8 may reliably estimate the trailer angle γ based on an end-to-end approach for angle estimation utilizing only the existing backup camera (e.g. imaging device 14 of the vehicle 12). As discussed herein, the deep neural networks provided for the detection of the trailer angle γ may contain thousands or millions of tunable parameters. Based on these parameters, the system 8 may accurately represent highly non-linear models while being very robust to noise. In some embodiments, convolutional layers may be trained to detect diverse features, similarly to human perception as such networks are capable of generalizing scenarios for later detection.

A drawback related to the implementation of deep convolution neural networks may include the labor-intensive involvement from human operators. For example, in order to provide feedback to the network, a human operator may be required to label and review thousands of samples to ensure accurate learning and operation of the network. Accordingly, in addition to providing the application of neural networks to identify the trailer angle γ, the disclosure also provides for methods of programming and training the neural networks discussed herein. These and other aspects of the disclosure are further detailed in the following description.

Referring still to FIGS. 1, 2, and 3, the vehicle 12 is embodied as a pickup truck that is pivotally attached to one embodiment of the trailer 10. The trailer 10 may comprise a box frame 16 with an enclosed cargo area 18. An axle 20 of the trailer may be operably coupled to wheels 22 and 24, and a tongue 26 may extend longitudinally forward from the enclosed cargo area 18. The illustrated trailer 10 comprises a trailer hitch connector in the form of a coupler assembly 28. The coupler assembly 28 is connected to a vehicle 12 via a hitch ball 30, which may be connected to the vehicle 12 by a drawbar. In operation, the coupler assembly 28 may latch onto the hitch ball 30 to provide a pivoting hitch point 32 that allows for articulation of the trailer angle γ between the vehicle 12 and the trailer 10.

As shown in FIG. 3, the trailer angle γ is shown in relation to a number of parameters of the vehicle 12 and the trailer 10. In operation, the kinematic model depicted in FIG. 3 may be utilized as the basis for the system 8 to control the navigation of the vehicle 12 to direct the trailer 10 along a calculated path. During such operations, the system 8 may monitor the trailer angle γ to ensure that the trailer 10 is accurately guided by the vehicle 12. The parameter that may be utilized for the model include, but are not limited to, the following:

δ: steering angle at steered wheels 40 of the vehicle 12;

α: yaw angle of the vehicle 12;

β: yaw angle of the trailer 10;

γ: trailer angle between the vehicle 12 and the trailer 10 (γ=β−α);

W: wheelbase length between a front axle 42 and a rear axle 44 of the vehicle 12;

L: drawbar length between the hitch point 32 and the rear axle 44 of the vehicle 12; and

D: trailer length between the hitch point 32 and axle 20 of the trailer 10 or effective axle for multiple axle trailers.

It should be appreciated that additional embodiments of the trailer 10 may alternatively couple with the vehicle 12 to provide a pivoting connection, such as by connecting with a fifth wheel connector. It is also contemplated that additional embodiments of the trailer 10 may include more than one axle and may have various shapes and sizes configured for different loads and items, such as a boat trailer or a flatbed trailer.

In some embodiments, the trailer backup assist system 8 may also include the imaging device 14 located at the rear of the vehicle 12 and configured to image a rear-vehicle scene. The imaging device 14 may be centrally located at an upper region of a vehicle tailgate 46 such that the imaging device 14 is elevated relative to the tongue 26 of the trailer 10. The imaging device 14 has a field of view 48 located and oriented to capture one or more images that may include the tongue 26 of the trailer 10 and the hitch ball 30, among other things. Captured images or image data may be supplied to a controller 50 of the trailer backup assist system 8. As discussed herein, the image data may be processed by the controller 50 to determine the trailer angle γ between the vehicle 12 and the trailer 10.

Referring now to FIG. 2, the controller 50 may comprise a microprocessor 52 and/or other analog and/or digital circuitry for processing one or more logic routines stored in a memory 54. The logic routines may include one or more trailer angle detection routines 56, which may comprise one or more deep learning neural networks as well as operating routines 58, which may be configured to guide the vehicle 12. Information from the imaging device 14 or other components of the trailer backup assist system 8 may be supplied to the controller 50 via a communication network of the vehicle 12, which can include a controller area network (CAN), a local interconnect network (LIN), or other conventional protocols used in the automotive industry. It should be appreciated that the controller 50 may be a stand-alone dedicated controller or may be a shared controller integrated with the imaging device 14 or other component of the trailer backup assist system 8 in addition to any other conceivable onboard or off-board vehicle control systems.

In an exemplary embodiment, the controller 50 of the trailer backup assist system 8 may be configured to communicate with a variety of vehicle equipment. The trailer backup assist system 8 may include a vehicle sensor module 60 that monitors certain dynamics of the vehicle 12. The vehicle sensor module 60 may generate a plurality of signals that are communicated to the controller 50 and may include a vehicle speed signal generated by a speed sensor 62 and a vehicle yaw rate signal generated by a yaw rate sensor 64. A steering input device 66 may be provided to enable a driver to control or otherwise modify the desired curvature of the backing path of the trailer 10.

The steering input device 66 may be communicatively coupled to the controller 50 in a wired or wireless manner. In this configuration, steering input device 66 may provide the controller 50 with information defining the desired curvature of the backing path of the trailer 10. In response, the controller 50 may process the information and generate corresponding steering commands that are supplied to a power assist steering system 68 of the vehicle 12. In some embodiments, the steering input device 66 may comprise a rotatable knob 70 operable to rotate to positions that may correspond to an incremental change to the desired curvature of a backing path of the trailer 10.

According to some embodiments, the controller 50 of the trailer backup assist system 8 may control the power assist steering system 68 of the vehicle 12 to operate the steered wheels 40 to direct the vehicle 12 in such a manner that the trailer 10 reacts in accordance with the desired curvature of the backing path of the trailer 10. The power assist steering system 68 may be an electric power-assisted steering (EPAS) system that includes an electric steering motor 74 for turning the steered wheels 40 to a steering angle δ based on a steering command generated by the controller 50. In this configuration, the steering angle δ may be sensed by a steering angle sensor 76 of the power assist steering system 68 and provided to the controller 50. The steering command may be provided for autonomously steering the vehicle 12 during a backup maneuver and may alternatively be provided manually via a rotational position (e.g., a steering wheel angle) of the steering input device 66 or the rotatable knob 70.

In some embodiments, the steering input device 66 (e.g. steering wheel) of the vehicle 12 may be mechanically coupled with the steered wheels 40 of the vehicle 12, such that the steering input device 66 may move in concert with steered wheels 40 via an internal torque, thereby preventing manual intervention with the steering input device 66 during autonomous steering of the vehicle 12. In such instances, the power assist steering system 68 may include a torque sensor 80 that senses torque (e.g., gripping and/or turning) on the steering input device 66, which may not be expected from autonomous control of the steering input device 66. Such unexpected torque may be detected by the controller 50 to indicate manual intervention by the driver. In some embodiments, external torque applied to the steering input device 66 may serve as a signal to the controller 50 that the driver has taken manual control and for the trailer backup assist system 8 to discontinue autonomous steering functionality.

The controller 50 of the trailer backup assist system 8 may also communicate with a vehicle brake control system 82 of the vehicle 12 to receive vehicle speed information, such as individual wheel speeds of the vehicle 12. Additionally or alternatively, vehicle speed information may be provided to the controller 50 by a powertrain control system 84 and/or the speed sensor 62, among other conceivable means. It is conceivable that individual wheel speeds may be used to determine a vehicle yaw rate, which can be provided to the controller 50, in the alternative or in addition, to the vehicle yaw rate measured by the yaw rate sensor 64 of the vehicle sensor module 60. In some embodiments, the controller 50 may provide braking commands to the vehicle brake control system 82, thereby allowing the trailer backup assist system 8 to regulate the speed of the vehicle 12 during a backup maneuver of the trailer 10. It should be appreciated that the controller 50 may additionally or alternatively regulate the speed of the vehicle 12 via interaction with the powertrain control system 84.

Through interaction with the power assist steering system 68, the vehicle brake control system 82, and/or the powertrain control system 84 of the vehicle 12, the potential for unacceptable trailer backup conditions can be reduced. Examples of unacceptable trailer backup conditions include, but are not limited to, a vehicle over-speed condition, a high trailer angle rate, trailer angle dynamic instability, a trailer jackknife condition, sensor failure, and the like. In such circumstances, the driver may be unaware of the failure until the unacceptable trailer backup condition is imminent or already happening. In order to avoid such conditions, the controller 50 may be configured to accurately monitor the trailer angle γ thereby providing feedback to ensure accurate operation.

According to some embodiments, the controller 50 may communicate with one or more devices, including a vehicle alert system 86, which may prompt visual, auditory, and tactile warnings. For instance, vehicle brake lights 88 and vehicle emergency flashers may provide a visual alert and a vehicle horn 90 and/or speaker 92 may provide an audible alert. Additionally, the controller 50 and/or vehicle alert system 86 may communicate with a human machine interface (HMI) 82 of the vehicle 12. The HMI 82 may include a touchscreen vehicle display 96, such as a center-stack mounted navigation or entertainment display capable of displaying images indicating the alert. Such an embodiment may be desirable to notify the driver of the vehicle 12 that an unacceptable trailer backup condition is occurring. Further, it is contemplated that the controller 50 may communicate via wireless communication with one or more electronic portable devices, such as portable electronic device 98, which is shown embodied as a smartphone. The portable electronic device 98 may include a display for displaying one or more images and other information to a user. In response, the portable electronic device 98 may provide feedback information, such as visual, audible, and tactile alerts.

Referring now to FIGS. 4A and 4B, an exemplary process diagram of the trailer angle detection routine is shown. In some implementations, trailer angle detection routine may be taught or otherwise programmed based on image or video data captured by the imaging device 14. This image data may then be processed by an image processing apparatus 104 (e.g. a high-performance image processing computer) to identify the trailer angle γ in the image data captured by the imaging device 14. Once the image data is labeled identifying the trailer angle γ, the labeled image data may be supplied to a neural network 102 for training. The neural network 102 may then process the image data and apply the labeled image data to independently identify the trailer angle γ in image data that is not labeled. Accordingly, once the neural network 102 is trained such that the trailer angle γ can successfully be identified by the hitch angle detection routine within an acceptable or predetermined level of error, the system 8 may be configured to utilize the neural network 102 to detect the trailer angle γ by utilizing unlabeled image data captured by the imaging device 14.

In reference to the FIGS. 4A and 4B, the hitch angle detection routine 56 will be described in reference to a training process 56a and an operating process 56b. The training process 56a may utilize one or more image processing techniques to identify the trailer angle γ in each image or frame of the image data. For example, the image processing apparatus 104 may process the image data via a Hough transform to identify lines extending from the coupler assembly 28 in order to identify a directional vector of the trailer 10. Similarly, the image processing apparatus 104 may process the image data by a convolution-based technique and various edge detection and filtering processes. Once the trailer angle γ is identified in the image data, the image data may be labeled with the trailer angle γ and provided as training inputs to the neural network 102.

The image processing apparatus 104 may be implemented as a component of the system 8 and/or as a separate computerized system. In some implementations, the processing requirements and steps necessary to accomplish the training process 56a may be more rigorous and/or demanding than the processing steps typically required from the controller 50. Accordingly, the processing steps for the training process 56a may be completed by a system comprising increased processing capacity or processing power for image processing to limit an expense of the dedicated hardware provided for the controller 50 in each system 8. For example, the controller 50 utilized for the training process 56a may be configured to have more capable image processing engines, processors capable of increased processing speeds, and generally more advanced system architecture than the controller 50 utilized for the operating process 56b. In this way, the system 8 may be effectively designed to promote economical manufacturing of the system 8 for training embodiments as well as operating embodiments, which may be incorporated in consumer products (e.g. the vehicle 12).

Referring first to the training process 56a, the image processing apparatus 104 may first receive image data from the imaging device 14 depicting the coupler assembly 28. The image data may be supplied sequentially or stored in a memory device 106 and supplied to the image processing apparatus 104 in bulk. Based on the image data, the image processing apparatus 104 may be configured to supply information to the neural network 102 such that the neural network 102 learns how to accurately detect the trailer angle γ. As further discussed herein, the image processing apparatus 104 may further manipulate the image data and monitor the accuracy of the neural network 102 to accommodate a self-supervised deep learning technique embodied by the training process 56a. In this way, the training process 56a may be configured to train the neural network 102 without requiring specialty hardware (e.g. a hitch angle detection sensor) in connection with the vehicle 12. In this way, the image data may be captured by any vehicle in connection with a trailer and supplied to provide custom training of the neural network 102 and/or improve the training of the neural network 102.

In operation, the image processing apparatus 104 may process the image data via a plurality of processing modules. For example, in some implementations, the image processing apparatus 104 may comprise a pre-processing module 104a. The pre-processing module 104a may be configured to crop each image frame captured by the imaging device 14. The cropping of the image data may be consistently processed based on the positional relationship of the hitch ball 30 in the field of view 48. For example, the hitch ball 30 may be registered or identified in the image data for each image set and corresponding hitch ball 30 for the vehicle 12. Once identified, the controller 50 may be configured to crop the image data from the imaging device 14 based on predetermined extents or a portion of the image data designated in relation to the location of the hitch ball 30 in the field of view 48. In this way, the image data supplied to the neural network 102 may be limited in positional variation to improve the sample data provided to train the neural network 102.

In some embodiments, the controller 50 may further be configured to process the data via an image augmentation module 104b. The image augmentation module 104b may be configured to augment the image data by a variety of techniques. For example, the cropped image data received from the image pre-processing module 104a may be augmented by the image augmentation module 104b by various techniques including, but not limited to, flipping, rotating, translating, scaling, color enhancing, histogram stretching, noise filtering, selective noise inclusion, etc. Following processing of the image data via the image pre-processing module 104a and/or the image augmentation module 104b, the controller 50 may further process the image data via one or more processing techniques or feature extraction techniques (e.g. Hough transform) to identify the trailer angle γ and label each frame of the image data via a trailer angle labeling module 104c. In this configuration, the trailer angle labeling module 104c may identify the trailer angle γ to be implemented as training data attributed to each of the frames of the image data that may be input into the neural network 102. In this way, the training process 56a may provide for the image data from the imaging device 14 to be processed and input into the neural network 102 with the trailer angle γ of each image frame identified in order to train the parameters of the neural network 102 to accurately identify the trailer angle γ from only the image data.

Once the image data is received by the neural network 102, a deep learning procedure may be implemented to regress or estimate the trailer angle γ. For example, the neural network 102 may be implemented as a deep convolutional network. The architecture of the neural network 102 may be a plurality of convolutional networks followed by activation functions. To help avoiding overfitting, dropout layers and other regularization techniques may be implemented. In an exemplary embodiment, fully connected layers at the end of the neural network 102 are responsible identifying that outputting the trailer angle γ. Since the object of the neural network 102 may be to perform a regression task, an activation function may not be utilized at the output.

In general, the neural network 102 may comprise a plurality of neurons 110, which may be arranged in a three-dimensional array comprising a width, a depth, and a height. The arrangement of the neurons 110 in this configuration may provide for each layer (e.g. dimensional cross-section of the array) to be connected to a small portion of the preceding layer. In this way, the network 102 may process the data through regression to reduce each image to a single vector to identify the trailer angle γ. Accordingly, the neural network 102 may transform each frame of the image data layer by layer from original pixel values to the final output. In general, the specific architecture of the neural network 102 may vary and as may be understood by those having ordinary skill in the art, the training process 56a may begin with a pre-trained model. In this way, the training process 56a may be utilized to fine-tune the pre-trained, convolutional neural network 102 to accurately detect the trailer angle γ from the image data captured by the imaging device 14. Examples of pre-trained models that may be implemented for the training process 56a may include, but are not limited to, the following: LeNet, AlexNet, ZF Net, GoogLeNet, VGGNet, ResNet, etc.

In order to optimize the training process 56a, the image data supplied to the neural network 102 may be selected to improve the accuracy of the identification of the trailer angle γ. Accordingly, in some implementations, the image processing apparatus 104 may selectively supply the labeled image data to the neural network 102 based on one or more predetermined parameters. The predetermined parameters utilized to select the image data that is supplied to the neural network 102 may generally be configured to provide image data to the neural network 102 that conforms to a plurality of training requirements. For example, the training requirements may include, but are not limited to a quantity of image samples, a trailer type and environment diversity of the image samples, an angle diversity of the image samples (e.g. variation in trailer angle γ), and an angle distribution of the image data (e.g. an even distribution of occurrences of the trailer angle γ in the image data). Accordingly, the image processing apparatus may select the image frames of the labeled image data such that the image data supplied to the neural network 102 conforms to the training requirements.

The quantity of image samples selected by the image processing apparatus 104 may correspond to an initial requirement necessary to satisfy the diversity and distribution requirements. That is, the quantity of image frames should be numerous enough to depict a variety of trailer types, environmental conditions, and lighting conditions in the image data. As discussed later in reference to FIG. 5, the appearance and corresponding identification of the trailer 10 may widely vary based on the conditions in which the image data is captured. Additionally, the images or image frames selected by the image processing apparatus 104 may vary in trailer angle over a desired range of angles. The requirements for the image data may further include that the trailer angle γ is balanced and evenly distributed such that each trailer angle γ is represented a similar number of times in the image data. In this way, the image processing apparatus 104 may select the image data to represent a variety of trailer types with approximately even distribution of each angle in an operating range of the system 8.

While the selection of the training of the neural network 102 may be accomplished based solely on the image data in some instances, the image frames captured by the imaging device 14 may not include sufficient variety or samples to satisfy the preferred data set supplied to the training process 56a. For example, the image data may not include a sufficient variety of examples of the trailer angle γ, variations in the type of the trailer 10, and/or variations in lighting or environmental conditions. Such limitations in the image data may be related to the typical operation of the vehicle 12.

For example, typical sets of image data collected by the imaging device 14 may not include samples of various angles of the trailer angle γ. Such limitations in the image data may be caused by limited variation in the trailer angle γ during common operation. That is, the variation in the trailer angle γ may be less than 20 degrees throughout the vast majority of operation of the system 8. Accordingly, the image processing apparatus 104 may further be configured to generate image frames and/or augment the image data in existing image frames to supplement the image data supplied to the neural network 102. For example, if a trailer angle is missing from the image data captured by the imaging device 14, the image processing apparatus 104 may augment the existing image data and/or generate simulated image data. The augmentation and simulation of the image data is further discussed in reference to FIG. 6.

Referring now to FIG. 5B, once the neural network 102 is trained, the operating process 56b of the hitch angle identification routine 56 may be processed without the need of the trailer angle γ identified by the angle labeling module 104c of the image processing apparatus 104. Accordingly, the operation of the operating process 56b may be streamlined to require less processing power than the training process 56a. For example, the controller 50 may similarly process the image data via a pre-processing module 50a. The pre-processing module 50a may be configured to crop each image frame received from the imaging device 14 based on the positional relationship of the hitch ball 30 in the field of view 48. Next, the controller 50 may process the cropped image data via an image enhancement module 50b.

The result of the steps completed by the pre-processing module 50a and/or the enhancement module 50b may generate normal image data that may be more uniform than the image data received directly from the imaging device 14. Such uniformity in proportions in relation to the hitch-ball 30 and image quality (e.g. contrast, noise, etc.) may provide for the trailer angle identification routine 56 to improve the successful identification of the trailer angle γ in the image data without requiring the computationally intensive steps required to identify the trailer angle γ via the angle labeling module 104c of the image processing apparatus 104 as described in the training process 56a. Accordingly, the disclosure may provide for improved operation and accuracy of the system 8 based on both the image processing steps provided by the modules 50a, 50b and the utilization of the neural network 102.

As previously discussed, the neural network 102 may be configured to receive the image data from the pre-processing module 50a and/or the enhancement module 50b. Upon completion of the analysis of each frame of the image data, the controller 50 may output the corresponding trailer angle γ. Accordingly, the systems and methods described herein may provide for the training and physical implementation of the system 8 that may effectively train the neural network 102 such that the trailer angle γ may be accurately identified in a variety of environmental conditions, lighting conditions, and for a variety of trailer topographies and orientations.

FIGS. 5A, 5B, and 5C demonstrate image samples of a variety of trailer types, trailer angles, and environmental or lighting conditions. A first set 120 of the image samples demonstrates a plurality of trailer hitch configurations. A second set 122 of the image samples demonstrates a plurality of lighting or environmental conditions for a plurality of trailer hitch configurations. Finally, a third set 124 of the image samples demonstrates a plurality of hitch angles for a trailer hitch configuration. As demonstrated by the image samples, each image sample may vary in appearance based on a significant variety of factors that may change the processing characteristics required for the neural network 102 to accurately identify the trailer angles γ. For this reason, the disclosure may provide for the neural network 102 to be provided a variety of image samples to support the recognition of the trailer angles γ in a variety of conditions.

As shown, each of the images shown in FIGS. 5A, 5B, and 5C are cropped such that the hitch of the coupler assembly 28 is consistently demonstrated in reference to a position of the hitch-ball 30. During the training process 56a and the operating process 56b, the pre-processing modules 104a and 50a may normalize the proportion of the image data such that the neural network 102 may be provided uniform image data to improve the identification of the trailer angle γ. The uniformity may further be adjusted via the image augmentation module 104b for the training process 56a and the image enhancement module 50b for the operating process 56b. The processing steps for each of the image augmentation module 104b and the image enhancement module 50b may be similar in reference to the processing of the images for uniformity. Such uniformity in proportions in relation to the hitch-ball 30 and image quality (e.g. contrast, noise, etc.) may provide for the trailer angle identification routine 56 to improve the successful identification of the trailer angle γ in the image data without requiring the secondary measurements from mechanical detection devices.

As previously discussed, the image processing apparatus 104 may be implemented in the vehicle 12 in some implementations; however, the image processing apparatus 104 may comprise more sophisticated image processing hardware configured to identify the trailer angle γ in the image data for training the neural network 102 via the training process 56a. Once trained, the neural network 102 may operate to identify the trailer angle γ directly from the image data, and, as such, the sophistication and related expense of the controller 50 including the image enhancement module 50b may be diminished relative to the image processing apparatus 104. Additionally, as previously discussed, the training process 56a may be completed by a device separate from the vehicle 12 to limit the expense of the controller 50. Accordingly, the neural network 102 may provide for improved economy while maintaining accurate identification of the trailer angle γ.

Referring now to FIG. 6, a flow chart of the training process 56a is shown providing further details the selection of the image data for training the neural network 102. The process 56a may begin by capturing image data with the imaging device demonstrating a field of view 48 including the coupler assembly 28 (130). Though discussed in reference to a single imaging device (e.g. the imaging device 14), the image data may be gathered by a number of imaging devices and reported to a central server via a wireless communication network and/or by accessing the image data stored in one or more memory devices incorporated in vehicles 12. In this way, the image data may be collected effectively to represent a variety of environments and operating conditions.

Once the image data is captured, the process 56a or method may continue by pre-processing the image data via the pre-processing module 104a (132). The pre-processing module 104a may be configured to crop each image frame captured by the imaging device 14. The cropping of the image data may be consistently processed based on the positional relationship of the hitch ball 30 in the field of view 48. For example, the hitch ball 30 may be registered or identified in the image data for each image set and corresponding hitch ball configuration 30 for the vehicle 12. Once identified, the controller 50 may be configured to crop the image data from the imaging device 14 based on predetermined extents or a portion of the image data designated in relation to the location of the hitch ball 30 in the field of view 48. In this way, the image data supplied to the neural network 102 may be limited in positional variation to improve the sample data provided to train the neural network 102.

Next, in step 134, the image augmentation module 104b may augment the image data by a variety of techniques (134). For example, the cropped image data received from the image pre-processing module 104a may be augmented by various techniques including, but not limited to, flipping, rotating, translating, scaling, color enhancing, histogram stretching, noise filtering, selective noise inclusion, etc. Following processing of the image data via the image pre-processing module 104a and/or the image augmentation module 104b, the controller 50 may further process the image data via one or more digital image processing or feature recognition techniques (e.g. Hough transform) to identify the trailer angle γ and label each frame of the image data via a trailer angle labeling module 104c (136). In this configuration, the trailer angle labeling module 104c may identify the trailer angle γ to be implemented as training data attributed to each of the frames of the image data that may be input into the neural network 102.

In step 138, the image processing apparatus 104 may be configured to compare the image set captured by the imaging device 14 with a variety of parameters that may be required for a complete set of teaching data. As previously discussed, the parameters for the image data may require a variety and distribution of trailer angles, trailer types, and/or environmental or lighting conditions. To address this requirement, the image processing apparatus 104 may select images representing each of a plurality values of the trailer angle γ. For example, the image processing apparatus 104 may select images including a range of angles specified for training (e.g. a range 100 deg. to +100 deg. from a centerline for the trailer angle γ). Additionally, the image processing apparatus 104 may select the images to include a precision of (for example 0.01 degree). Accordingly, the complete data set of images for training the neural network 102 may comprise 20,000 images to include the complete range of values for the trailer angle γ and at the desired resolution between the angles demonstrated.

In step 140, the process 56a may continue to identify whether each of image samples in the range of trailer angles γ is available in the image data. If each image in the desired range of angles is included image data, the method may continue to step 142 to supply the image data to the neural network 102 for training. Similarly, the image processing apparatus 104 may supply images or frames of the image data to the neural network 102 that comply with the requirements of the data set to complete the training process 56a. If the image data is incomplete as determined in step 140, the image processing apparatus 104 continues to generate the missing image data. For example, the image processing apparatus 104 may augment or adjust the image data with the image augmentation module 104b to adjust the appearance of the trailer to complete the image set by updating the trailer angle γ, changing the trailer type, and/or adjusting environmental or lighting conditions (144).

For example, in step 138, the image processing apparatus 104 may compare the image data to the desired data set to determine if the data set includes a depiction of the trailer 10 at a requested angle θ (e.g. θ=85.23 degrees). In step 140, the image processing apparatus 104 may identify that the requested angle θ is not included in the image data. In response to identifying that the image data does not include the requested angle θ, the image processing apparatus 104 may access one or more image frames that is close to the requested angle θ and adjust the appearance of the image frame to match the requested angle θ as further discussed in reference to FIG. 7.

Referring now to FIGS. 7A, 7B, and 7C, a plurality of stored image frames 150 are shown. Continuing with the previous example, in response to the image processing apparatus 104 identifying that the requested angle θ (e.g. θ=85.23 degrees) is not included in the image data, as labeled by the trailer angle labeling module 104c, the image processing apparatus 104 may identify the stored image frames 150 that are closest to the requested angle θ. The image processing apparatus 104 may then compare the available angles ϕ_i of the stored image frames 150 to determine if the available angles ϕ_i are less than a predetermined angular change δ from the requested angle θ. (e.g. |θ−ϕ_i|<δ). If the difference between the requested angle θ and the available angles ϕ_i (e.g. ϕ₁ and ϕ₂ as shown in FIGS. 7A and 7B) is less than the predetermined angular difference δ, the image processing apparatus 104 may rotate all or a portion of one of the stored images such that a new or augmented image 152 or frame is created. Once the augmented image 152 is created, the image processing apparatus 104 may supply the augmented image 152 to the neural network 102 for training. In this way, the image processing apparatus 104 may identify and generate the image data necessary to fulfill the angular range requirements necessary to accurately train the neural network 102.

Referring now to FIGS. 8A and 8B, the image processing apparatus 104 may similarly augment the image data to adjust the trailer type, the lighting, and/or the environment depicted in the image data. For example, a stored image 160 may be augmented by the image processing apparatus 104 to generate an augmented image 162. In this example, the augmented image 162 may be processed to adjust an appearance of a background 164 and the coupler assembly 28 depicted in the stored image 160. In this way, the image processing apparatus 104 may be configured to generate the augmented image 162 to supplement and complete the image data captured by the imaging device 14.

It is to be understood that variations and modifications can be made on the aforementioned structures and methods without departing from the concepts of the present invention, and further it is to be understood that such concepts are intended to be covered by the following claims unless these claims by their language expressly state otherwise.

It will be understood by one having ordinary skill in the art that construction of the described device and other components is not limited to any specific material. Other exemplary embodiments of the device disclosed herein may be formed from a wide variety of materials, unless described otherwise herein.

For purposes of this disclosure, the term “coupled” (in all of its forms, couple, coupling, coupled, etc.) generally means the joining of two components (electrical or mechanical) directly or indirectly to one another. Such joining may be stationary in nature or movable in nature. Such joining may be achieved with the two components (electrical or mechanical) and any additional intermediate members being integrally formed as a single unitary body with one another or with the two components. Such joining may be permanent in nature or may be removable or releasable in nature unless otherwise stated.

It is also important to note that the construction and arrangement of the elements of the device as shown in the exemplary embodiments is illustrative only. Although only a few embodiments of the present innovations have been described in detail in this disclosure, those skilled in the art who review this disclosure will readily appreciate that many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.) without materially departing from the novel teachings and advantages of the subject matter recited. For example, elements shown as integrally formed may be constructed of multiple parts or elements shown as multiple parts may be integrally formed, the operation of the interfaces may be reversed or otherwise varied, the length or width of the structures and/or members or connector or other elements of the system may be varied, the nature or number of adjustment positions provided between the elements may be varied. It should be noted that the elements and/or assemblies of the system may be constructed from any of a wide variety of materials that provide sufficient strength or durability, in any of a wide variety of colors, textures, and combinations. Accordingly, all such modifications are intended to be included within the scope of the present innovations. Other substitutions, modifications, changes, and omissions may be made in the design, operating conditions, and arrangement of the desired and other exemplary embodiments without departing from the spirit of the present innovations.

It will be understood that any described processes or steps within described processes may be combined with other disclosed processes or steps to form structures within the scope of the present device. The exemplary structures and processes disclosed herein are for illustrative purposes and are not to be construed as limiting.

It is also to be understood that variations and modifications can be made on the aforementioned structures and methods without departing from the concepts of the present device, and further it is to be understood that such concepts are intended to be covered by the following claims unless these claims by their language expressly state otherwise.

The above description is considered that of the illustrated embodiments only. Modifications of the device will occur to those skilled in the art and to those who make or use the device. Therefore, it is understood that the embodiments shown in the drawings and described above is merely for illustrative purposes and not intended to limit the scope of the device, which is defined by the following claims as interpreted according to the principles of patent law, including the Doctrine of Equivalents.

Claims

1. A method for training an image-based trailer identification system comprising: capturing a plurality of captured images in a field of view;identifying a detected trailer angle for a trailer in connection with a vehicle in each of the captured images;comparing the captured images and the corresponding trailer angles to a predetermined image set comprising a plurality of teaching trailer angles;identifying at least one required trailer angle of the teaching trailer angles that is not included in the captured images;generating a simulated angle image based on the captured images, wherein the simulated image comprises a depiction of the trailer in connection with the vehicle at the at least one required angle not included in the captured images; andsupplying the simulated angle image to the identification system for training.

2. The method according to claim 1, wherein the identifying of the detected trailer angle is determined by digitally processing the image data via a feature extraction technique.

3. The method according to claim 2, wherein the feature extraction technique comprises at least one of a line detection.

4. The method according to claim 1, wherein the identification system comprises a neural network.

5. The method according to claim 1, wherein the plurality of teaching trailer angles comprises a range of trailer angles for identification.

6. The method according to claim 5, wherein the plurality of teaching trailer angles comprises an increment between each trailer angle in the range of trailer angles.

7. The method according to claim 1, further comprising: detecting a lighting condition in the captured images and comparing the lighting condition to a required lighting condition for training the identification system.

8. The method according to claim 7, further comprising: in response to the lighting condition of the captured images differing from the required lighting condition, adjusting the lighting of at least one of the captured images to generate the simulated lighting image.

9. The method according to claim 8, further comprising: supplying the simulated lighting image to the identification system for training.

10. The method according to claim 1, wherein the image data is captured by a plurality of vehicles, each configured to capture the images in the field of view.

11. The method according to claim 10, further comprising: storing the captured images from the plurality of vehicles and compiling the captured images for comparing the captured images.

12. A system for training a neural network comprising: at least one imaging device configured to capture a plurality of captured images of a trailer in connection with a vehicle;an image processing apparatus configured to:identify a detected trailer angle from each of the captured images based on an image processing routine;compare the captured images and the corresponding detected trailer angles to a predetermined image set comprising a plurality of teaching trailer angles;identify at least one required trailer angle of the teaching trailer angles that is not included in the captured images;generate a simulated angle image based on the captured images; andsupply the simulated angle image to the neural network for training.

13. The system according to claim 12, wherein the simulated angle image comprises a depiction of the trailer in connection with the vehicle at the at least one required angle not included in the captured images.

14. The system according to claim 12, wherein the image processing routine comprises a digital image processing technique comprising a feature extraction technique.

15. The system according to claim 14, wherein the feature extraction technique comprises at least one of a line detection.

16. The system according to claim 12, wherein the plurality of teaching trailer angles comprises a range of trailer angles for identification.

17. The system according to claim 16, wherein the plurality of teaching trailer angles comprises an increment between each trailer angle in the range of trailer angles.

18. The system according to claim 12, wherein the image processing apparatus is further configured to: detect a lighting condition in the captured images and compare the lighting condition to a required lighting condition for training the identification system.

19. The system according to claim 18, wherein the image processing apparatus is further configured to: in response to the lighting condition of the captured images differing from the required lighting condition, adjust the lighting to at least one of the captured images and generate the simulated lighting image.

20. A system for training a neural network comprising: at least one imaging device configured to capture a plurality of captured images of a trailer in connection with a coupler of a vehicle;an image processing apparatus configured to: identify a first trailer type in the captured images based on a feature extraction technique configured to identify at least one feature of the trailer relative to the coupler;compare the first trailer type to a plurality of teaching trailer types;generate a simulated trailer image comprising a second trailer type of the plurality of teaching trailer types that is not included in the captured images; andsupply the simulated trailer image to the neural network for training.