An experience learning method and device for automatic driving of a vehicle

Abstract

The application discloses an experience learning method and device for automatic driving of a vehicle, and comprises the following steps: obtaining sensor data through a vehicle perception system; obtaining surrounding vehicle motion data through the sensor data; analyzing the surrounding vehicle motion data to obtain an optimal learning object vehicle; taking target vehicle motion data corresponding to the optimal learning object vehicle as a training set to perform model training, and obtaining an abnormal driving behavior mode; updating a vehicle automatic driving model according to the abnormal driving behavior mode; and generating a real-time driving strategy based on the updated vehicle automatic driving model. The automatic driving vehicle can learn the behaviors of surrounding vehicles to improve its own behaviors, so that the automatic driving vehicle can learn abnormal events, and the adaptability and robustness of the automatic driving vehicle itself are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to an experience learning method and device for vehicle automatic driving. BACKGROUND

[0002] At present, most of the automatic driving vehicles rely on limited training data before deployment to learn driving behavior based on machine learning and deep learning technology. Due to the great differences in road rules, traffic conditions, pedestrian behavior, etc. in different regions, the driverless vehicle relying only on past driving data may be difficult to accurately identify and understand various traffic conditions and make correct decisions. In addition, the driverless vehicle is easily affected by heavy rain, snow and other adverse weather conditions and other surrounding vehicles, which affects the performance of the sensors of the driverless vehicle, thereby reducing the robustness and performance of the automatic driving system. SUMMARY

[0003] The technical problem to be solved by the present application is to provide an experience learning method and device for vehicle automatic driving to improve the adaptability of the automatic driving vehicle itself.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is:

[0005] An experience learning method for vehicle automatic driving, comprising:

[0006] obtaining sensor data through a vehicle perception system;

[0007] obtaining surrounding vehicle motion data through the sensor data;

[0008] analyzing the surrounding vehicle motion data to obtain a best learning object vehicle;

[0009] training a model with target vehicle motion data corresponding to the best learning object vehicle as a training set to obtain an abnormal driving behavior pattern;

[0010] updating a vehicle automatic driving model according to the abnormal driving behavior model;

[0011] generating a real-time driving strategy based on the updated vehicle automatic driving model.

[0012] To solve the above technical problems, another technical scheme adopted by the present application is:

[0013] An experience learning device for vehicle automatic driving, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to realize each step of the above-mentioned experience learning method for vehicle automatic driving.

[0014] The present application has the beneficial effect that: through obtaining sensor data by a vehicle perception system and further obtaining surrounding vehicle motion data, the best learning object vehicle is obtained based on the analysis of the surrounding vehicle motion data, and then the target vehicle motion data of the best learning object vehicle is learned to update the vehicle automatic driving model and generate real-time driving strategy, so that the automatic driving vehicle can learn the behavior of surrounding vehicles to improve its own behavior, thereby enabling the automatic driving vehicle to learn abnormal events and improve its own adaptability and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A step flow chart of an experience learning method for vehicle automatic driving in an embodiment of the present application;

[0016] Figure 2 Another step flow chart of an experience learning method for vehicle automatic driving in an embodiment of the present application;

[0017] Figure 3 A step flow chart of obtaining surrounding vehicle motion data in an experience learning method for vehicle automatic driving in an embodiment of the present application;

[0018] Figure 4 A schematic diagram of detecting the best learning object vehicle in an experience learning method for vehicle automatic driving in an embodiment of the present application;

[0019] Figure 5 A structure schematic diagram of an experience learning device for vehicle automatic driving in an embodiment of the present application. DETAILED DESCRIPTION

[0020] To explain the technical content, achieved purposes and effects of the present application in detail, the following will be described in combination with embodiments and the accompanying drawings.

[0021] In the prior art, when an automatic driving vehicle encounters heavy snow weather or is blocked by a large vehicle, the automatic driving vehicle will be difficult to accurately identify the lane line, resulting in reduced reliability and safety of the automatic driving vehicle. In response to these challenges, further research and improvement are needed on the design and training strategy of the automatic driving system.

[0022] Please refer to Figure 1 An experience learning method for vehicle automatic driving, comprising:

[0023] Obtaining sensor data through a vehicle perception system;

[0024] Obtaining surrounding vehicle motion data through the sensor data;

[0025] analyze the surrounding vehicle motion data to obtain an optimal learning object vehicle;

[0026] train a model using target vehicle motion data corresponding to the optimal learning object vehicle as a training set to obtain an abnormal driving behavior pattern;

[0027] update a vehicle automatic driving model according to the abnormal driving behavior model;

[0028] generate a real-time driving strategy based on the updated vehicle automatic driving model.

[0029] As can be seen from the above description, the present application has the beneficial effects that: after obtaining sensor data by a vehicle perception system and further obtaining surrounding vehicle motion data, an optimal learning object vehicle is obtained based on analysis of the surrounding vehicle motion data, and then the target vehicle motion data of the optimal learning object vehicle is learned to update a vehicle automatic driving model and generate a real-time driving strategy, so that the automatic driving vehicle can learn the behavior of surrounding vehicles to improve its own behavior, thereby enabling the automatic driving vehicle to learn abnormal events and improve its own adaptability and robustness.

[0030] Further, the sensor data includes image data and radar data;

[0031] The surrounding vehicle motion data obtained from the sensor data includes:

[0032] The image data is feature-extracted by a residual layer of a neural network to obtain image features;

[0033] The radar data is feature-extracted by a recurrent layer of a neural network to obtain radar features;

[0034] The image features and the radar features are fused by an attention mechanism of a neural network to obtain fused features;

[0035] The fused features are processed by a fully connected layer of a neural network to obtain the surrounding vehicle motion data.

[0036] As can be seen from the above description, after the image data and the radar data are respectively feature-extracted by a residual layer and a recurrent layer of a neural network, the image features and the radar features are fused by an attention mechanism, and then processed by a fully connected layer, so that more accurate surrounding vehicle motion data can be obtained.

[0037] Further, the fusion of the image features and the radar features by the attention mechanism to obtain fused features includes:

[0038] The attention weight parameters are learned by a backpropagation algorithm during model training;

[0039] multiply the image features and the radar features with corresponding elements in the set of attention weight parameters respectively to obtain feature values;

[0040] perform softmax normalization on the feature values to obtain attention weights corresponding to the image features and the radar features respectively;

[0041] perform weighted summation on the image features and the radar features through the attention weights to obtain the fusion features.

[0042] As can be seen from the above description, the attention weight parameters learned through the back propagation algorithm during the model training process can obtain the corresponding weight elements and attention weights of the image features and the radar features, which can reflect the weight proportion of the image features and the radar features, and the final fusion features can be obtained based on the product of the attention weights and the image features and the radar features.

[0043] Further, the feature extraction on the image data through the residual layer of the neural network includes:

[0044] performing cropping, scaling and grayscale processing on the image data to obtain a standard image;

[0045] performing feature extraction on the standard image through the residual layer of the neural network to obtain the image features.

[0046] As can be seen from the above description, the image data is generated into a unified standard image through the cropping, scaling and grayscale processing, which improves the efficiency of subsequent feature extraction on the image data.

[0047] Further, the radar data includes laser radar data and millimeter wave radar data;

[0048] The feature extraction on the radar data through the recurrent layer of the neural network includes:

[0049] performing data conversion and normalization on the laser radar data and the millimeter wave radar data to obtain standard radar data;

[0050] performing feature extraction on the standard radar data through the recurrent layer of the neural network to obtain the radar features.

[0051] As can be seen from the above description, the laser radar data and the millimeter wave radar data are uniformly converted into standard radar data, which improves the efficiency of subsequent feature extraction on the radar data.

[0052] Further, before obtaining the surrounding vehicle motion data through the sensor data, it includes:

[0053] determining whether the perception system is functioning properly, and if not, controlling the vehicle to slow down and stop and activating a backup system;

[0054] if so, performing the step of obtaining surrounding vehicle motion data from the sensor data.

[0055] As can be seen from the above description, by controlling the vehicle to slow down and stop and activating a backup system when it is detected that the perception system is not functioning properly, the driving function and safety of the vehicle can be maintained based on the backup system.

[0056] Further, before the step of performing model training on the target vehicle motion data corresponding to the optimal learning object vehicle as a training set, the method further comprises:

[0057] obtaining self-vehicle motion data;

[0058] determining whether the target vehicle motion data is better than the self-vehicle motion data, and if so, using the target vehicle motion data as the training set.

[0059] As can be seen from the above description, by comparing the target vehicle motion data with the self-vehicle motion data, learning is only performed when the target vehicle motion data is better than the self-vehicle motion data, thereby avoiding learning of non-optimal vehicle motion data.

[0060] Further, the step of analyzing the surrounding vehicle motion data to obtain an optimal learning object vehicle comprises:

[0061] identifying and classifying all vehicles within the detection range by a target detection algorithm to obtain target vehicles and their vehicle types;

[0062] evaluating the stability of all the target vehicles by a machine learning algorithm, and selecting the target vehicle with the highest stability as the optimal learning object vehicle.

[0063] As can be seen from the above description, based on the target detection algorithm, all vehicles within the detection range and their corresponding vehicle types can be accurately identified, and then the target vehicle with the highest stability is selected as the optimal learning object vehicle by the machine learning algorithm, thereby realizing identification of the optimal learning object vehicle.

[0064] Further, before the step of evaluating the stability of all the target vehicles by a machine learning algorithm, the method further comprises:

[0065] determining whether the vehicle type of the target vehicle is the same as the vehicle type of the host vehicle, and if not, performing feature processing on the vehicle type of the target vehicle according to the self-vehicle type.

[0066] From the above description, when the vehicle model of the target vehicle is different from the vehicle model of the host vehicle, the vehicle model of the target vehicle is processed by the vehicle model of the host vehicle, so that the vehicle motion data of the target vehicle can be applied to the autonomous vehicle.

[0067] Another embodiment of the present application provides a vehicle automatic driving experience learning device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to realize each step of the vehicle automatic driving experience learning method.

[0068] The vehicle automatic driving experience learning method and device provided by the present application can be applied to the vehicle automatic driving scene, realize learning of the autonomous vehicle on abnormal events, and improve the adaptability and robustness of the autonomous vehicle itself, which will be described below through a specific embodiment.

[0069] Embodiment one

[0070] Please refer to Figure 1 and Figure 2 A vehicle automatic driving experience learning method comprises the following steps:

[0071] S1, obtaining sensor data through a vehicle perception system; for example, obtaining image data through a camera, obtaining laser radar data and millimeter wave radar data through a laser radar and a millimeter wave radar respectively, or obtaining required data through other sensors; the laser radar can measure the distance and shape of an object, and the millimeter wave radar is used to detect the position and speed of a target object.

[0072] S02, judging whether the perception system is normally running, if not, controlling the vehicle to slow down and stop and starting a backup system; if yes, performing S2.

[0073] S2, obtaining surrounding vehicle motion data through sensor data, specifically:

[0074] Please refer to Figure 3 , the convolutional neural network can use different models for feature extraction, and in an optional implementation, a residual network structure (ResNet50) is used, comprising the following steps:

[0075] S21, pre-processing the collected data, for example, performing image data processing including cropping, scaling and grayscale processing to obtain standard images; and performing data conversion and normalization on laser radar data and millimeter wave radar data to obtain standard radar data; the input of the model can be represented as a combination of image data and radar data;

[0076] S22, feature extraction is performed on the standard image through a residual layer (ResNet) of the neural network to obtain image features; that is, the ResNet is used to process the image collected by the automobile camera and extract features in the image; for example, the input of the image captured by the camera of the self-driving automobile is a tensor Image_input, the dimension of which is [width, height, channels], wherein width and height represent the width and height of the image respectively, and channels represent the number of channels of the image; through the image processing branch (ResNet) of the neural network, the feature map Featrue_img of the final image is obtained based on operations such as convolution (CONC), pooling (Pool), activation function (ReLu), and batch normalization (BatchNorm), that is, the image features are obtained; wherein, IDBlock represents an Identity block, which is a standard block in the ResNet; Conv Block represents a convolution block;

[0077] S23, feature extraction is performed on the standard radar data through a recurrent layer (RNN) of the neural network and changes in the surrounding environment are recorded to obtain radar features, that is, the spatiotemporal dynamic features in the radar data such as target motion trajectory and speed change; the RNN is used to process the time series information of the radar data, such as the change of the radar point in time; for example, the radar data can be represented as a tensor Rader_input, the dimension of which is [num_radar_points, radar_features], wherein num_radar_points represents the number of radar points, and radar_features represents the number of features of each radar point; through the operation of the radar branch of the neural network based on the recurrent layer, the final radar data feature is obtained and represented as Featrue_radar, that is, the radar features are obtained;

[0078] S24, the image features and the radar features are fused through a multimodal attention mechanism (Multimodal Attention Mechanism) of the neural network to obtain fused features, in particular:

[0079] S241, the image features (Featrue_img) and the radar features (Featrue_radar) are flattened into vectors to obtain an image feature vector (Flattend_featrue_img) and a radar feature vector (Flattend_featrue_radar) respectively, and then the image feature vector and the radar feature vector are connected together to obtain a concatenated feature vector (Concatenated_featrues);

[0080] S242, the attention weight parameter set is learned by a back propagation algorithm during model training; for example, an attention weight parameter vector Wa is used as the distribution weight in the attention mechanism, and the length of the parameter vector Wa is the same as the length of the connection feature vector; that is, the parameter vector Wa is used as the attention weight parameter set, and the parameter vector Wa is learned by a back propagation algorithm during model training;

[0081] S243, the image features and the radar features are multiplied by corresponding elements in the attention weight parameter set to obtain feature values; that is, each element in the connection feature vector is multiplied by the corresponding element in the parameter vector Wa to obtain the corresponding feature value;

[0082] S244, the feature values are subjected to softmax normalization to obtain the attention weights corresponding to the image features and the radar features, respectively;

[0083] S245, the image features and the radar features are weighted and summed by the attention weights to obtain the fused features; that is, the connection feature vector is weighted and summed by the attention weights, and finally the fused feature vector (Attention_fused_featrues) is obtained.

[0084] S25, the fused features are processed by a fully connected layer (FC) of a neural network to obtain surrounding vehicle motion data; that is, the fused feature vector (Attention_fused_featrues) is input into the fully connected layer for processing, and the final output is represented as output, output = [x, y, vx, vy, angle]; wherein, the network output prediction result includes the position (x, y), velocity (vx, vy), direction (angle) and other motion data of the surrounding vehicle, and the motion data required to be predicted can be increased according to the demand, such as acceleration data; further, the perspective and blind spot of other vehicles in the vicinity can also be estimated using a deep learning algorithm, and a bird's eye view of the surrounding environment is generated in combination with the surrounding vehicle motion data; for example, the environment around the vehicle is projected onto a plane using perspective transformation and other technologies, thereby obtaining a bird's eye view; the bird's eye view can help the autonomous vehicle to detect obstacles, such as other vehicles or pedestrians. At the same time, the system also analyzes the obtained surrounding vehicle motion data in real time and checks whether the data is abnormal, and if it is abnormal, the surrounding vehicle motion data is regenerated; for example, when the sensor data is abnormal due to the influence of strong rain or heavy snow and other bad weather on the autonomous vehicle, the abnormal data needs to be detected and identified, and measures are taken to collect data again to ensure the accuracy and reliability of the data.

[0085] S3, analyze the surrounding vehicle motion data to obtain the best learning object vehicle, specifically:

[0086] Please refer to Figure 4 , S31, all vehicles in the detection range are identified and classified by a target detection algorithm to obtain target vehicles and their models; in an optional embodiment, when it is judged that the model of the target vehicle is different from the model of the host vehicle, the model of the target vehicle is processed according to the model of the host vehicle; that is, the feature conversion is needed to adapt to the learning needs of the host vehicle due to the size and characteristic differences between different models; including standardizing or normalizing the features to make the features have similar scales and eliminate the size differences between different models; on the converted data, a suitable reinforcement learning algorithm can be used for training to achieve effective learning and decision-making for sudden situations;

[0087] S32, the stability of all target vehicles is evaluated by a machine learning algorithm, and the target vehicle with the highest stability is selected as the best learning object vehicle; for example, the motion data such as vehicle position (x, y), speed (vx, vy), direction (angle) are used for evaluation;

[0088] S4, the target vehicle motion data corresponding to the best learning object vehicle is used as a training set for model training to obtain abnormal driving behavior patterns; the target vehicle motion data is used as a training set only when the host vehicle motion data is better than the target vehicle motion data; for example, a pre-defined model is used to detect the driving behavior of the autonomous vehicle, including bad acceleration, speed, braking, steering, unreasonable vehicle distance, violation of traffic rules, etc. If the behavior is found to be poor in some cases, the system determines that the correct behavior of surrounding vehicles needs to be learned to improve its own behavior; that is, when the autonomous vehicle judges that its driving behavior is unstable, it can improve its driving behavior by actively starting the learning process; due to the limited training data of autonomous vehicles, they cannot fully adapt to all driving situations; before the autonomous vehicle makes the next action, the vehicle's expected driving behavior is generated, that is, the best driving strategy, and the current best driving strategy may not be better than the surrounding vehicles; at the same time, the driving behavior of the autonomous vehicle is usually relative to other surrounding vehicles; if the driving behavior of the host vehicle is better than that of the surrounding vehicles, the current strategy can be considered as the best; otherwise, it is not good. By comparing the driving strategy of the surrounding vehicles in the current environment with the best strategy or the current driving behavior of the host vehicle, the autonomous vehicle can determine whether to update the best strategy or the current driving behavior;

[0089] S5, updating the vehicle automatic driving model according to the abnormal driving behavior model; for example, using a deep learning algorithm to train the collected driving data of the surrounding vehicles, and learning the driving behavior pattern of the best surrounding vehicle by minimizing the loss function.

[0090] S6, generating real-time driving strategy based on the updated vehicle automatic driving model; that is, the autonomous vehicle combines the vehicle's own planning algorithm (such as path planning and trajectory tracking algorithm) to formulate the corresponding driving strategy and action, and feeds back to the automatic driving model in real time to update the state.

[0091] For example, scenario one: when the autonomous vehicle encounters congestion, the conventional best strategy is to maintain a sufficient safety distance and yield to other vehicles to change lanes; however, this strategy may cause other vehicles to frequently cut in, thereby slowing down the autonomous vehicle's travel speed; in this case, the autonomous vehicle can optimize its driving strategy by collecting the behavior data of surrounding vehicles; for example, under the premise of ensuring safety, the autonomous vehicle can appropriately shorten the distance from the front vehicle, reduce the frequency of yielding to other vehicles to change lanes, and speed up the travel speed as much as possible to minimize the time loss caused by congestion; or adjust the speed, acceleration, etc. Parameters to drive more smoothly and reduce the number of braking and acceleration as much as possible, thereby improving ride comfort and fuel efficiency.

[0092] At the same time, the autonomous driving system will monitor abnormal events, such as encountering traffic situations that cannot be understood, emergencies, etc.; these abnormal situations may cause the system to be unable to respond or handle correctly, and therefore are considered as learning opportunities; when unable to understand the traffic situation or encounter an emergency, the autonomous driving system first takes appropriate action to ensure driving safety, and learns in real time during this process to optimize the driving strategy;

[0093] For example, scenario two: when driving on a highway in snowy weather, the autonomous vehicle drives to a turn and finds that the snowplow has cleared the snow to the sides of the road, and the snow melts into water that easily flows to the low side of the curve to form an icy road surface; if the current vehicle is driving on the lower lane of the curve, the vehicle is prone to skidding, and the system cannot understand the current situation; at this time, the autonomous vehicle will slow down to cope with the current complex road situation; then, the system will collect driving data and environmental data of surrounding vehicles, including the behavior of surrounding vehicles in similar situations and changes in road conditions; by analyzing these data and using machine learning algorithms, the system finds that vehicles driving on the high side of the curve have more stable driving behavior and can maintain a relatively fast speed; after obtaining this information, the autonomous vehicle will learn the driving behavior of these vehicles and include it in the consideration range of the driving strategy; so that when encountering similar situations, the autonomous vehicle will directly choose the high side of the curve as the driving road, thereby improving driving safety and efficiency.

[0094] Embodiment Two

[0095] Please refer to Figure 5 The application further provides an experience learning device for automatic driving of a vehicle, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and each step of the experience learning method for automatic driving of a vehicle according to Embodiment One is implemented when the processor executes the computer program.

[0096] In conclusion, the experience learning method and device for automatic driving of a vehicle provided by the application can realize that the automatic driving vehicle can learn the behavior of surrounding vehicles to improve its own behavior, so that the automatic driving vehicle can learn abnormal events and improve its own adaptability and robustness, by obtaining sensor data and further obtaining surrounding vehicle motion data by a vehicle perception system, obtaining the optimal learning object vehicle based on analysis of the surrounding vehicle motion data, learning the target vehicle motion data of the optimal learning object vehicle, updating the vehicle automatic driving model and generating a real-time driving strategy.

[0097] The above description is only an embodiment of the application, and does not limit the patent scope of the application, and any equivalent transformation or direct or indirect application in the related technical field based on the content of the specification and drawings of the application is also included in the patent protection scope of the application.

Claims

1. An experience learning method for automatic driving of a vehicle, characterized by, The method comprises the following steps: obtaining sensor data through a vehicle perception system; obtaining surrounding vehicle motion data through the sensor data; analyzing the surrounding vehicle motion data to obtain an optimal learning object vehicle; training a model with target vehicle motion data corresponding to the optimal learning object vehicle as a training set to obtain an abnormal driving behavior model; updating a vehicle automatic driving model according to the abnormal driving behavior model; generating a real-time driving strategy based on the updated vehicle automatic driving model; the analysis of the surrounding vehicle motion data to obtain the optimal learning object vehicle comprises: identifying and classifying all vehicles in a detection range through a target detection algorithm to obtain target vehicles and their vehicle types; evaluating the stability of all target vehicles through a machine learning algorithm, and selecting the target vehicle with the highest stability as the optimal learning object vehicle; the evaluation of the stability of all target vehicles through the machine learning algorithm comprises: determining whether the vehicle type of the target vehicle is the same as that of the host vehicle, and if not, performing feature processing on the vehicle type of the target vehicle according to the vehicle type of the host vehicle. 2.The experience learning method of automatic driving of a vehicle according to claim 1, characterized in that, The sensor data includes image data and radar data. The surrounding vehicle motion data obtained through the sensor data comprises: extracting features of the image data through a residual layer of a neural network to obtain image features; extracting features of the radar data through a recurrent layer of a neural network to obtain radar features; fusing the image features and the radar features through an attention mechanism of a neural network to obtain fused features; processing the fused features through a fully connected layer of a neural network to obtain the surrounding vehicle motion data. 3.The experience learning method of automatic driving of a vehicle according to claim 2, characterized in that, The fusion of the image features and the radar features through the attention mechanism comprises: learning a set of attention weight parameters through a back propagation algorithm during model training; multiplying the image features and the radar features by corresponding elements in the set of attention weight parameters to obtain feature values; performing softmax normalization on the feature values to obtain attention weights corresponding to the image features and the radar features, respectively; performing weighted summation on the image features and the radar features through the attention weights to obtain the fused features. 4.The experience learning method of automatic driving of a vehicle according to claim 2, characterized in that, The extraction of the image features through the residual layer of the neural network comprises: performing standard image processing including cropping, scaling, and grayscale processing on the image data to obtain a standard image; extracting features of the standard image through the residual layer of the neural network to obtain the image features. 5.The experience learning method of automatic driving of a vehicle according to claim 2, characterized in that, The radar data includes lidar data and millimeter wave radar data. The extraction of the radar features through the recurrent layer of the neural network comprises: performing data conversion and normalization on the lidar data and the millimeter wave radar data to obtain standard radar data; extracting features of the standard radar data through the recurrent layer of the neural network to obtain the radar features. 6.The experience learning method of automatic driving of a vehicle according to claim 1, wherein The method further comprises the following steps before obtaining the surrounding vehicle motion data through the sensor data: determining whether the perception system is functioning properly, and if not, controlling the vehicle to slow down and stop and activate a backup system; if so, performing the step of obtaining surrounding vehicle motion data from the sensor data.

7. The experience learning method for automatic driving of a vehicle according to claim 1, characterized by, The method further comprises, before the step of training a model using the target vehicle motion data corresponding to the optimal learning object vehicle as a training set, the steps of: obtaining self vehicle motion data; determining whether the target vehicle motion data is better than the self vehicle motion data, and if so, using the target vehicle motion data as the training set.

8. An experience learning device for automatic driving of a vehicle, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor, when executing the computer program, implements each step of the experience learning method for automatic driving of a vehicle according to any one of claims 1-7.

Images