Vehicle environment sensing method, device and equipment and storage medium

Abstract

The invention relates to the technical field of vehicle perception, and discloses a vehicle environment perception method, device and equipment and a storage medium, and the method comprises the steps: optimizing a class-level quaternary adversarial network model according to the historical surrounding environment data of a vehicle, and obtaining an environment perception model; performing classification and identification according to the environment sensing model and real-time surrounding environment data of the vehicle to obtain a prediction probability corresponding to each category of the detection target; environment sensing is carried out according to the prediction probability corresponding to each type of the detection target and the current natural environment condition of the vehicle, and environment sensing information is obtained; and executing a corresponding driving decision according to the environmental perception information. The sensing capability of the automobile under different environment conditions can be effectively improved, and the sensing accuracy is enhanced.

Description

Technical Field

[0001] The present application relates to the field of vehicle perception technology, and in particular to vehicle environment perception methods, devices, equipment and storage media. Background Art

[0002] The development of autonomous vehicles is inseparable from the accurate perception of the surrounding environment. The environmental perception system is the core component of autonomous vehicles. It collects information about the vehicle's surroundings through various sensors (such as radar, lidar, cameras, etc.) and processes this information into data that the vehicle can "understand" for navigation, obstacle avoidance, decision-making, etc. With the advancement of technology, environmental perception systems have been able to handle some basic traffic scenarios, such as lane detection, vehicle and pedestrian recognition, etc. However, existing environmental perception systems perform well under standardized test conditions, but their performance is often affected when faced with the complex and changing real world. For example, in bad weather (such as fog, rain, snow) or lighting conditions (such as at night or in tunnels), the sensor's perception ability will be limited, resulting in inaccurate data collection, which in turn affects the stability and reliability of the entire system. In addition, there are differences in road conditions, traffic rules, and driving behaviors in different regions, all of which put higher demands on environmental perception systems.

[0003] In order to improve the performance of environmental perception systems, machine learning and deep learning technologies are widely used in this field due to their powerful data processing and pattern recognition capabilities. Through training, these models can identify and classify different objects, predict the behavior of other vehicles and pedestrians, and thus improve the safety and efficiency of autonomous vehicles. Although machine learning and deep learning technologies have made certain progress in environmental perception, existing technologies still face some challenges. First, deep learning models usually require a large amount of labeled data for training, which may be difficult to obtain in certain specific scenarios. Secondly, the knowledge obtained by the model under specific environmental conditions is difficult to transfer to other environments, which limits the generalization ability of the model. In addition, the training and deployment process of the model is often time-consuming and costly, which may not be feasible in a rapidly changing environment. It can be seen that the environmental perception system of autonomous vehicles has limitations in quickly adapting to changing environmental conditions and lacks the necessary flexibility and scalability to adapt to diverse application scenarios.

[0004] Therefore, how to effectively improve the perception ability of automobiles under different environmental conditions and enhance perception accuracy is a problem that needs to be solved urgently.

[0005] The above contents are only used to assist in understanding the technical solution of the present application and do not constitute an admission that the above contents are prior art. Summary of the invention

[0006] The main purpose of this application is to provide a vehicle environment perception method, device, equipment and storage medium, aiming to solve the technical problem of how to effectively improve the perception ability of a car under different environmental conditions and enhance the perception accuracy.

[0007] To achieve the above objectives, the present application proposes a vehicle environment perception method, the method comprising:

[0008] According to the historical surrounding environment data of the vehicle, the class-level quaternary adversarial network model is optimized to obtain the environment perception model;

[0009] Classify and identify the environment according to the environment perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target;

[0010] Performing environmental perception according to the predicted probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information;

[0011] Execute corresponding driving decisions according to the environmental perception information.

[0012] In one embodiment, the optimizing the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle to obtain the environment perception model includes:

[0013] Acquire historical surrounding environment data of the vehicle, and perform a preprocessing operation on the historical surrounding environment data to obtain preprocessed historical surrounding environment data, wherein the preprocessing operation at least includes data cleaning, data standardization, and noise removal operations;

[0014] Dividing and labeling the preprocessed historical surrounding environment data to obtain source domain data and target domain data, wherein the source domain data is composed of source domain labeled samples, and the target domain data is composed of target domain labeled samples and target domain unlabeled samples;

[0015] The class-level four-element adversarial network model is optimized according to the source domain data and the target domain data to obtain an environmental perception model, wherein the class-level four-element adversarial network model includes a feature extractor, a discriminator, and a classifier.

[0016] In one embodiment, optimizing the class-level quaternary adversarial network model according to the source domain data and the target domain data to obtain an environment perception model includes:

[0017] Training the class-level quaternary adversarial network model according to the source domain data and the target domain data to obtain model performance parameters;

[0018] The adversarial network model is optimized according to the model performance parameters to obtain an environmental perception model.

[0019] In one embodiment, the class-level four-element adversarial network model is trained according to the source domain data and the target domain data to obtain model performance parameters, including:

[0020] The feature extractor, the discriminator and the classifier are trained according to the source domain data and the target domain data respectively to obtain the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category heart alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled sample;

[0021] Determine a feature extractor parameter, a discriminator parameter, and a classifier parameter according to at least one of the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-class center alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the prediction probability of the target domain unlabeled sample;

[0022] Model performance parameters are determined based on the feature extractor parameters, the discriminator parameters, and the classifier parameters.

[0023] In one embodiment, the feature extractor, the discriminator, and the classifier are trained respectively according to the source domain data and the target domain data to obtain the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category heart alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the prediction probability of the target domain unlabeled sample, including:

[0024] The feature extractor and the classifier are trained by using the source domain label samples and the target domain label samples to obtain source domain features, prediction results of source domain label samples, first target domain features, and prediction results of target domain label samples;

[0025] Performing feature extraction and classification recognition on the target domain unlabeled samples by using the trained feature extractor and the trained classifier to obtain a second target domain feature and a prediction result of the target domain unlabeled samples;

[0026] According to the source domain features, the prediction results of the source domain labeled samples, the first target domain features, the prediction results of the target domain labeled samples, the second target domain features and the prediction results of the target domain unlabeled samples, the cross entropy classification loss of the source domain labeled samples, the cross entropy classification loss of the target domain labeled samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled samples are calculated.

[0027] In one embodiment, the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category center alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the target domain unlabeled sample prediction probability are calculated according to the source domain feature, the prediction result of the source domain label sample, the first target domain feature, the prediction result of the target domain label sample, the second target domain feature, and the prediction result of the target domain unlabeled sample, including:

[0028] Calculating according to the source domain features, the prediction results of the source domain label samples, the first target domain features, and the prediction results of the target domain label samples to obtain the cross entropy classification loss of the source domain label samples and the cross entropy classification loss of the target domain label samples;

[0029] Determine the source domain data class center, the target domain data class center, and the combined data class center according to the number of source domain labeled samples, the number of target domain labeled samples, the source domain features, the first target domain features, the second target domain features, and the prediction results of the target domain unlabeled samples;

[0030] Determine the three-category center alignment loss according to the source domain data center, the target domain data center, and the combined data center;

[0031] Constructing a four-dimensional data pair according to the source domain feature, the source domain data class center, the first target domain feature, and the target domain data class center;

[0032] The discriminator judges the four-dimensional data pair to generate a domain prediction result;

[0033] Determining a class-level quadruple adversarial loss based on the domain prediction result and the quadruple data pair;

[0034] An entropy regularization loss of the prediction probability of the target domain unlabeled samples is determined according to the number of the target domain unlabeled samples and the prediction results of the target domain unlabeled samples.

[0035] In one embodiment, determining the feature extractor parameters, the discriminator parameters, and the classifier parameters according to at least one of the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category center alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the target domain unlabeled sample prediction probability includes:

[0036] Determine a feature extractor parameter according to the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-class heart alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the target domain unlabeled sample prediction probability;

[0037] Determine the discriminator parameters according to the three types of heart alignment losses;

[0038] The classifier parameters are determined according to the cross entropy classification loss of the source domain labeled samples, the cross entropy classification loss of the target domain labeled samples, and the entropy regularization loss of the predicted probability of the target domain unlabeled samples.

[0039] In addition, to achieve the above purpose, the present application also proposes a vehicle environment perception device, the vehicle environment perception device comprising:

[0040] An optimization module, used to obtain historical surrounding environment data of the vehicle, and optimize the class-level quaternary adversarial network model according to the historical surrounding environment data to obtain an environment perception model;

[0041] A prediction module is used to perform classification and identification based on the environmental perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target;

[0042] A perception module, used to perform environmental perception according to the predicted probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information;

[0043] An execution module is used to execute corresponding driving decisions according to the environmental perception information.

[0044] In addition, to achieve the above-mentioned purpose, the present application also proposes a vehicle environment perception device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the computer program is configured to implement the steps of the vehicle environment perception method described above.

[0045] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the vehicle environment perception method described above are implemented.

[0046] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the vehicle environment perception method described above are implemented.

[0047] The present application provides a vehicle environment perception method, which first optimizes a class-level four-element adversarial network model according to the historical surrounding environment data of the vehicle to obtain an environment perception model; performs classification and identification according to the environment perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target; performs environment perception according to the prediction probability corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environment perception information; and executes corresponding driving decisions according to the environment perception information, which can effectively improve the perception ability of the car under different environmental conditions and enhance the perception accuracy.

[0048] In summary, the present application optimizes the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle, thereby effectively improving the adaptability and recognition ability of the model to environmental changes. The predicted probability corresponding to each category of the detection target is quickly and accurately determined through the environmental perception model and the real-time surrounding environment data of the vehicle, and then the environment is perceived according to the predicted probability and the current natural environment conditions of the vehicle. The perceived environment perception information can be accurately obtained and the corresponding driving decisions can be executed, which overcomes the technical defects of the environmental perception system of autonomous driving vehicles in quickly adapting to changing environmental conditions. It can effectively improve the perception ability of the vehicle under different environmental conditions and enhance the perception accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0051] Figure 1 A schematic diagram of a process flow provided for the first embodiment of the vehicle environment perception method of the present application;

[0052] Figure 2 A vehicle perception system framework diagram provided for the first embodiment of the vehicle environment perception method of the present application;

[0053] Figure 3 A schematic diagram of a flow chart provided for the second embodiment of the vehicle environment perception method of the present application;

[0054] Figure 4 A schematic diagram of the structure of a class-level quaternary adversarial network model provided in Example 2 of the vehicle environment perception method of the present application;

[0055] Figure 5This is a schematic diagram of the module structure of the vehicle environment perception device according to an embodiment of the present application;

[0056] Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the vehicle environment perception method of an embodiment of the present application.

[0057] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0058] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0059] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0060] The main solution of the embodiment of the present application is: optimize the class-level four-element adversarial network model according to the historical surrounding environment data of the vehicle to obtain an environmental perception model; perform classification and identification according to the environmental perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target; perform environmental perception according to the prediction probability corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information; and execute corresponding driving decisions according to the environmental perception information.

[0061] In order to improve the performance of environmental perception systems, machine learning and deep learning technologies are widely used in this field due to their powerful data processing and pattern recognition capabilities. Through training, these models can identify and classify different objects, predict the behavior of other vehicles and pedestrians, and thus improve the safety and efficiency of autonomous vehicles. Although machine learning and deep learning technologies have made certain progress in environmental perception, existing technologies still face some challenges. First, deep learning models usually require a large amount of labeled data for training, which may be difficult to obtain in certain specific scenarios. Secondly, the knowledge obtained by the model under specific environmental conditions is difficult to transfer to other environments, which limits the generalization ability of the model. In addition, the training and deployment process of the model is often time-consuming and costly, which may not be feasible in a rapidly changing environment. It can be seen that the environmental perception system of autonomous vehicles has limitations in quickly adapting to changing environmental conditions and lacks the necessary flexibility and scalability to adapt to diverse application scenarios. Therefore, how to effectively improve the perception ability of vehicles under different environmental conditions and enhance the accuracy of perception is a problem that needs to be solved urgently.

[0062] This application optimizes the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle, effectively improving the model's adaptability and recognition ability to environmental changes, and quickly and accurately determines the predicted probability corresponding to each category of the detection target through the environmental perception model and the real-time surrounding environment data of the vehicle, and then performs environmental perception based on the predicted probability and the current natural environmental conditions of the vehicle. It can accurately obtain perceived environmental perception information and execute corresponding driving decisions, overcoming the technical defects of the environmental perception system of autonomous driving vehicles in quickly adapting to changing environmental conditions, and can effectively improve the vehicle's perception ability under different environmental conditions and enhance perception accuracy.

[0063] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a vehicle environment sensing device, etc. The following takes the vehicle environment sensing device as an example to illustrate this embodiment and the following embodiments.

[0064] Based on this, the embodiment of the present application provides a vehicle environment perception method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the vehicle environment perception method of the present application.

[0065] In this embodiment, the vehicle environment perception method includes steps S10 to S40:

[0066] Step S10, optimizing the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle to obtain an environmental perception model.

[0067] It should be noted that the surrounding environment data refers to the vehicle surrounding environment information collected by sensors such as images, radars, and lidars, including but not limited to road conditions, traffic signs, the positions and speeds of other vehicles and pedestrians. Historical surrounding environment data refers to various types of sensor data collected by the vehicle in different surrounding environments within a certain time range. With this data, the class-level quaternary adversarial network model can be trained and optimized to better adapt to various environmental changes.

[0068] It can be understood that the Class-Level Quaternary Adversarial Network (CQAN) model is a deep learning model specially designed to process cross-domain data. It learns the feature alignment between the source domain and the target domain through the construction of quadruple data, thereby improving the generalization ability of the model in different fields. In the perception system of autonomous vehicles, the Class-Level Quaternary Adversarial Network (CQAN) model can maintain good performance under different natural environmental conditions (such as different weather and lighting conditions, etc.).

[0069] In the specific implementation, the class-level quaternion adversarial network is used to conduct adversarial training based on the historical surrounding environment data of the vehicle, and then the optimal model performance parameters obtained from the adversarial training are transferred to adapt to the vehicle environment perception needs in different environments. Transfer learning is a machine learning method that allows the model to transfer the knowledge learned on one task to another related task. Through transfer learning, the model can use the knowledge learned in one domain (source domain) to improve the performance in another domain (target domain), even if the target domain has less data or is significantly different from the source domain. In the automotive field, transfer learning can be used in a variety of scenarios, such as autonomous driving, vehicle identification, traffic flow prediction, etc. For example, a model trained under clear weather conditions can be adapted to perception tasks in rainy days through transfer learning.

[0070] It is worth noting that the environmental perception model, i.e. the optimized class-level quaternary adversarial network, is formed by updating the parameters of the class-level quaternary adversarial network model according to the optimal model performance parameters obtained by adversarial training of the class-level quaternary adversarial network, so as to improve its perception ability in different environments. Figure 2 As shown, Figure 2 This is a framework diagram of the vehicle perception system. The vehicle perception system includes a sensor unit, a data preprocessing module, a class-level four-element adversarial network model, an environmental adaptability module, and an autonomous driving vehicle control system.

[0071] The class-level quaternary adversarial network in this embodiment combines adversarial learning and transfer learning to form an efficient learning framework. This framework addresses the problem that existing data cannot fully cover all driving environment scenarios, especially under changing weather conditions and lighting conditions, the accuracy of existing systems may be reduced. The class-level quaternary adversarial network makes full use of limited annotated data through a quaternary adversarial strategy to learn and transfer knowledge to other environmental perception tasks. This strategy strengthens the model's adaptability and recognition ability to environmental changes by aligning the differences in data features from different driving scenarios. This strategy effectively solves the problem of deep learning's dependence on a large amount of annotated data in environmental perception tasks, and improves the accuracy and robustness of the vehicle perception system when facing different environmental conditions. The class-level quaternary adversarial network model can not only reduce dependence on large-scale annotated data sets, but also maintain efficient and accurate perception performance in a changing traffic environment.

[0072] Step S20, classifying and identifying according to the environmental perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target.

[0073] It should be noted that real-time surrounding environment data refers to the surrounding environment information collected in real time by various sensors during the driving process of the vehicle, including but not limited to road conditions, traffic signs, the position and speed of other vehicles and pedestrians, etc. Classification and recognition refers to the use of environmental perception models to process these real-time data, identify the categories of detection targets, and calculate the corresponding prediction probabilities.

[0074] It is understandable that the detection targets can be various traffic participants, such as pedestrians, vehicles, lanes, obstacles, etc. The classification and recognition process involves inputting real-time surrounding environment data into the trained environmental perception model. The model will process the data through feature extractors and classifiers, and finally output the category of each detection target and its corresponding prediction probability.

[0075] In the specific implementation, the real-time surrounding environment data is standardized and enhanced through the preprocessing module to ensure the consistency of data quality and improve the robustness of the model. Preprocessing may include operations such as denoising, normalization, and data enhancement to adapt to the data format and distribution used in model training. The processed real-time surrounding environment data is fed into the feature extractor of the environmental perception model. The feature extractor is usually composed of multiple convolutional layers, pooling layers, and activation functions. Its purpose is to extract high-level features that are helpful for classification from the raw data. The extracted features are then passed to the classifier module. The classifier is usually composed of a fully connected layer and an output layer. The output layer uses a softmax function or other activation functions to calculate the predicted probability of each category, thereby realizing the classification and recognition of the detected target.

[0076] Step S30, performing environmental perception according to the predicted probabilities corresponding to the various categories of the detection targets and the current natural environment conditions in which the vehicle is located, to obtain environmental perception information.

[0077] It should be noted that the current natural environmental conditions of the vehicle refer to the various natural environmental factors that the vehicle faces during driving, such as weather conditions (sunny, rainy, foggy, etc.), lighting conditions (daytime, night, dusk, tunnels, etc.), temperature, humidity, etc. These environmental conditions have an important impact on the performance of the vehicle's perception system, especially in bad weather or low light conditions, the sensor's perception ability may be limited, thus affecting the accuracy and reliability of the environmental perception model. Therefore, the acquisition of environmental perception information needs to comprehensively consider the impact of these natural environmental conditions. The vehicle's environmental perception system will monitor and obtain the natural environmental conditions data of the vehicle in real time, which can be obtained through the vehicle's own sensors or external data sources. For example, the vehicle can obtain the current temperature through the built-in temperature sensor, obtain the lighting conditions through the photosensitive sensor, and detect whether it is raining through the rain sensor. In addition, the vehicle can also receive real-time weather forecasts and road conditions information from the meteorological station or traffic management center through the vehicle-mounted communication system. After obtaining the natural environmental conditions data of the vehicle, the environmental perception system will combine these data with the real-time surrounding environment data for comprehensive analysis to ensure that the environmental perception model can maintain high accuracy and robustness under different natural environmental conditions.

[0078] It can be understood that environmental perception information is obstacle information, collision information, traffic rules information, lane information, etc. determined based on the predicted probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle. This information helps the vehicle better understand the surrounding environment and make safer and more reasonable decisions.

[0079] It is worth noting that the environmental perception model will combine real-time surrounding environment data and current natural environmental conditions, and conduct a comprehensive analysis and evaluation of the vehicle's driving environment through a series of complex algorithms and calculation processes. For example, when driving in rainy days, the environmental perception model will pay special attention to the slipperiness of the road surface and the impact of raindrops on the sensor, thereby adjusting the judgment and prediction of road conditions. At night or in tunnels, the model will use enhanced image processing technology to improve the ability to recognize traffic signs and obstacles.

[0080] In the specific implementation, environmental perception information such as obstacle information, collision information, traffic rules information, lane information, etc. is determined according to the predicted probability corresponding to each category of the detection target and the current natural environment conditions of the vehicle. For example, according to the predicted probability of the detection target and the current natural environment conditions, the obstacles in the surrounding environment are identified and analyzed. The obstacles may include other vehicles, pedestrians, animals, stationary objects, etc. The vehicle perception system will evaluate the position, speed, direction of movement and potential collision risk of the obstacles, calculate the potential collision probability based on the obstacle information, and evaluate the collision risk level. Combined with the current natural environment conditions, traffic rules information such as traffic signs, signal lights and road markings are identified and interpreted. This helps vehicles comply with traffic regulations and ensure driving safety. For example, at a red light, the system will remind the driver to stop and wait; at a speed limit sign, the system will suggest the driver to reduce the speed. The lane lines, lane boundaries and lane changes of the road are analyzed to provide lane keeping and lane change suggestions for the vehicle. When the lane lines are blurred or unclear, the vehicle perception system uses enhanced image processing technology to improve the recognition ability of lane information and ensure that the vehicle drives safely in the lane.

[0081] Step S40: executing corresponding driving decisions according to the environmental perception information.

[0082] It should be noted that driving decision refers to driving behavior selection, including obstacle avoidance decision, lane change decision, overtaking decision, emergency braking decision and vehicle distance control decision, etc. This embodiment does not impose specific restrictions on this.

[0083] It is understandable that the corresponding driving decision is executed according to the environmental perception information. For example, when the environmental perception information is collision information, the system will immediately execute the emergency braking decision to avoid the accident. If the environmental perception information shows that a vehicle in front changes lanes and enters the driving lane of the vehicle, the system will execute the obstacle avoidance decision, slow down appropriately and adjust the driving direction to ensure a safe distance.

[0084] In the specific implementation, the vehicle's driving state is adjusted according to the executed driving decision, including but not limited to adjusting parameters such as vehicle speed, steering angle, acceleration, etc. For example, when executing a lane change decision, the system will calculate a safe lane change trajectory and gradually adjust the steering angle to ensure that the vehicle smoothly transitions to the new driving lane.

[0085] The present embodiment provides a vehicle environment perception method. The present embodiment obtains an environment perception model by first optimizing a class-level four-element adversarial network model according to historical surrounding environment data of the vehicle; performing classification and identification according to the environment perception model and the real-time surrounding environment data of the vehicle to obtain prediction probabilities corresponding to each category of the detection target; performing environment perception according to the prediction probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environment perception information; and executing corresponding driving decisions according to the environment perception information, which can effectively improve the perception ability of the automobile under different environmental conditions and enhance the perception accuracy.

[0086] In summary, this embodiment optimizes the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle, thereby effectively improving the adaptability and recognition ability of the model to environmental changes. The predicted probability corresponding to each category of the detection target is quickly and accurately determined through the environmental perception model and the real-time surrounding environment data of the vehicle, and then the environment is perceived according to the predicted probability and the current natural environment conditions of the vehicle. The perceived environment perception information can be accurately obtained and the corresponding driving decisions can be executed, which overcomes the technical defects of the environmental perception system of the autonomous driving vehicle in quickly adapting to changing environmental conditions, and can effectively improve the perception ability of the vehicle under different environmental conditions and enhance the perception accuracy.

[0087] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can refer to the above introduction, and will not be repeated later. Figure 3 , the step S10 further includes steps S101-S103:

[0088] Step S101, acquiring historical surrounding environment data of the vehicle, and performing a preprocessing operation on the historical surrounding environment data to obtain preprocessed historical surrounding environment data, wherein the preprocessing operation at least includes data cleaning, data standardization, and noise removal operations.

[0089] It should be noted that the acquisition of the vehicle's historical surrounding environment data can be achieved through the sensor unit of the vehicle's environmental perception system. The sensor unit includes but is not limited to cameras, radars, laser radars (LiDAR), etc., which can collect environmental data around the vehicle in real time.

[0090] It can be understood that preprocessing operations refer to a series of processing steps on raw data to ensure data quality and improve the accuracy of subsequent analysis, including data cleaning, data standardization, and noise removal operations. Data cleaning mainly removes invalid and erroneous data to ensure data quality; data standardization converts data into a unified format and range for subsequent processing; noise removal operations include removing redundant and irrelevant data to improve data processing efficiency.

[0091] Step S102, dividing and labeling the preprocessed historical surrounding environment data to obtain source domain data and target domain data, wherein the source domain data is composed of source domain labeled samples, and the target domain data is composed of target domain labeled samples and target domain unlabeled samples.

[0092] It should be noted that, in this embodiment, source domain data refers to vehicle surrounding environment data collected under specific conditions, which is used to train the class-level quaternary adversarial network model to identify and classify various environmental features. The source domain data consists of source domain label samples, which are source domain data samples with the "source domain" label. The target domain data refers to environmental data collected under different conditions, which consists of target domain label samples and target domain unlabeled samples, that is, target domain data samples with the "target domain" label and unlabeled target domain data samples, which are used for further training of the model and improvement of generalization ability.

[0093] Step S103, optimizing the class-level four-element adversarial network model according to the source domain data and the target domain data to obtain an environmental perception model, wherein the class-level four-element adversarial network model includes a feature extractor, a discriminator, and a classifier.

[0094] It should be noted that, let X s ∈R ns×ds represents the source domain data, X t ∈R nt×dt represents the target domain data, Υ s ∈R ns×1 represents the label of the source domain data, where R represents a real number set, d s and d t is the dimension of the two domain data, n s is the number of source domain samples, n t Represents the number of target domain sample data. The target domain data consists of two parts: t ={X tl ,X tu}, where X tl Represents a small amount of labeled data in the target domain, and its corresponding label is X tu Represents the unlabeled data of the target domain. The CQAN algorithm processes the features of the two domain data through multiple discriminators and realizes various types of confrontation by aligning various features. The label D of the same domain data pair s ∈R n×1 Set to all 0 vectors, labels D of different domain data pairs t ∈R n×1 Set to all 1 vector, where n=n s +n t .

[0095] It is understandable that the class-level quaternary adversarial network model is trained using source domain data, and the target domain data is used to fine-tune the model and improve its generalization ability to achieve an optimized class-level quaternary adversarial network model. During the training process, the model continuously optimizes the feature representation of the source domain data through the adversarial learning mechanism, while reducing the distribution difference between the source domain and target domain data to improve the adaptability and accuracy of the model in unseen environments.

[0096] It is worth noting that the class-level quaternary adversarial network model includes a feature extractor, a classifier, and a discriminator, wherein the feature extractor includes a source domain feature extractor and a target domain feature extractor. The feature extractor is responsible for extracting useful feature information from the input data, and the classifier classifies the data according to the extracted features. The discriminator is used to determine whether the input data belongs to the source domain or the target domain, while the adversarial network is trained against the feature extractor so that the features generated by the feature extractor have better consistency between the source domain and the target domain.

[0097] like Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of the class-level four-element adversarial network model. The class-level four-element adversarial network consists of three parts: feature extractor G s and G t , discriminator and shared classifier G C Since the feature dimensions of the source domain and the target domain are different, the source domain and the target domain have their own feature extractors G s and G t , two feature extractors first layer network G s1 and G t1 The parameters of the remaining network layer G st Shared parameters. s and target domain data X t Input to the feature extractor G s and G t , the first layer G of the two feature extractors s1 and G t1 The input nodes are d s and d t , which is the source domain data X s and target domain data X t The dimensions of the feature space in which they are located are the same as the output nodes of the first layer network. After the feature extractor, the data of both domains are reduced to d dimensions. Classifier G C It is shared across domains and is intended to output the classification prediction probabilities of the source domain and the target domain. There are a total of C categories of data, and there are C discriminators in the network, namely D c , c = 1,…,C. For each discriminator D c For example, its input is a four-dimensional data pair According to the discriminator Dc The output of can calculate the class-level four-element adversarial loss L D According to the feature extractor G s and G t The output of the classifier G C For the output of the source domain data and the output of the target domain data, three types of alignment losses L are calculated. cla_mean According to the classifier G C For the output of the source domain label sample and the label Υ of the source domain data s Calculate the cross entropy classification loss L of the source domain label sample s According to the classifier G C For the output of the target domain labeled sample and the target domain with a small number of labeled samples, L Calculate the cross entropy classification loss L of the target domain label sample tl , according to the classifier G C For the output of the unlabeled sample in the target domain, the entropy regularization loss L of the predicted probability of the unlabeled sample in the target domain is calculated entropy .

[0098] It is worth noting that the core of the class-level quadruple adversarial network lies in its unique quadruple data construction method, which guides the model to learn richer and more discriminative feature representations through the association between instances and class centroids. In the framework of adversarial learning, the model trains the discriminator by minimizing a specific loss function, which aims to distinguish the feature representations of the source domain and the target domain. The feature extractor is trained to generate features that can mislead the discriminator, thereby promoting the model to learn more generalized features under different environmental conditions. In this way, the class-level quadruple adversarial network not only reduces the dependence on large-scale annotated datasets, but also significantly improves the recognition and adaptation capabilities of the vehicle perception system in the face of environmental changes. This optimized network structure and loss function design provide a more powerful and flexible environmental perception solution for autonomous vehicles, enabling them to maintain efficient and accurate perception performance in a variety of traffic environments and conditions.

[0099] In a feasible implementation, step S103 may include: training the class-level four-element adversarial network model according to the source domain data and the target domain data to obtain model performance parameters; optimizing the adversarial network model according to the model performance parameters to obtain an environmental perception model.

[0100] It should be noted that training the class-level quaternary adversarial network model includes training the feature extractor, discriminator, and classifier separately. During the training process, the feature extractor and classifier are jointly trained using source domain data to ensure that the model can accurately identify and classify environmental features in the source domain data. Subsequently, the target domain data is introduced, and the parameters of the feature extractor are gradually adjusted through the synergy of the adversarial network and the discriminator, so that the performance of the model on the source domain and target domain data is more balanced, thereby improving the generalization ability of the model under different environmental conditions. The transfer learning technology is used to transfer the common features learned on the source domain data to the target domain data. Through transfer learning, the model can better utilize the rich information in the source domain data and accelerate the learning process on the target domain data, thereby achieving a higher recognition accuracy in a shorter time.

[0101] It can be understood that the model performance parameters include feature extractor parameters, discriminator parameters and classifier parameters, that is, the optimal solutions of feature extractor parameters, discriminator parameters and classifier parameters, wherein the feature extractor parameters include source domain feature extractor parameters and target domain feature extractor parameters. By updating the corresponding parameters in the class-level four-element adversarial network model according to the feature extractor parameters, discriminator parameters and classifier parameters, the class-level four-element adversarial network model can be optimized to obtain an environmental perception model.

[0102] In a feasible implementation manner, the class-level four-element adversarial network model is trained according to the source domain data and the target domain data to obtain model performance parameters, including: training the feature extractor, the discriminator and the classifier according to the source domain data and the target domain data respectively to obtain the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled samples; determining the feature extractor parameters, the discriminator parameters and the classifier parameters according to at least one of the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled samples; and determining the model performance parameters according to the feature extractor parameters, the discriminator parameters and the classifier parameters.

[0103] It should be noted that the cross entropy classification loss of source domain label samples is an indicator to measure the classification performance of feature extractors and classifiers on source domain data. By minimizing this loss function, the model can be ensured to have a high recognition accuracy on source domain data. The cross entropy classification loss of target domain label samples is used to measure the classification performance of feature extractors and classifiers on target domain data. By adjusting the parameters of the feature extractor, the classification loss on target domain data is minimized, thereby improving the generalization ability of the model on unseen data. The three-category alignment loss is used to ensure that the feature representations of each category in the source and target domain data have good alignment in the feature space. By minimizing this loss function, similar features in the source and target domain data can be closer in the feature space, thereby improving the recognition consistency of the model under different environmental conditions. The class-level four-element adversarial loss is the core of adversarial network training, which is used to ensure that the features generated by the feature extractor have better consistency between the source and target domains. Through adversarial training, the feature extractor and the discriminator compete with each other, making the features generated by the feature extractor difficult to be distinguished by the discriminator, thereby improving the generalization ability of the model under different environmental conditions. The entropy regularization loss of the prediction probability of unlabeled samples in the target domain is used to ensure that the prediction results of the model on the unlabeled data in the target domain have higher uncertainty. By minimizing this loss function, the prediction results of the model on the unlabeled data in the target domain can be made more diversified, thereby improving the generalization ability of the model when facing unknown data.

[0104] It can be understood that by jointly optimizing the cross entropy classification loss of source domain labeled samples, the cross entropy classification loss of target domain labeled samples, the three-category heart alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the prediction probability of target domain unlabeled samples to solve the optimal solution of feature extractor parameters, discriminator parameters and classifier parameters, and then updating the model parameters, we can obtain an environmental perception model with excellent comprehensive performance. This model not only has a high recognition accuracy on source domain data, but also shows good generalization ability on target domain data.

[0105] It is worth noting that by constructing four-dimensional data pairs, using the discriminator to learn the similarity measurement between data, and performing fine-grained classification and alignment of features through independently trained classifiers, it is possible to effectively capture the similarity between data and improve the effect of cross-domain feature alignment, so that the vehicle perception system can quickly learn and adapt to new environmental conditions, reduce dependence on large amounts of retraining data, improve data utilization efficiency, enhance perception accuracy, and have better flexibility and scalability, thereby enhancing the flexibility and scalability of the environmental perception system of autonomous vehicles, enabling them to better adapt to the complex and changeable actual driving environment.

[0106] In a feasible implementation manner, the feature extractor, the discriminator and the classifier are trained respectively according to the source domain data and the target domain data to obtain the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category heart alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled sample, including: training the feature extractor and the classifier with the source domain label sample and the target domain label sample to obtain the source domain feature, the prediction result of the source domain label sample, the first target domain feature and the prediction result of the target domain label sample; The trained feature extractor and the trained classifier perform feature extraction and classification recognition on the target domain unlabeled samples to obtain the second target domain features and the prediction results of the target domain unlabeled samples; the cross entropy classification loss of the source domain labeled samples, the cross entropy classification loss of the target domain labeled samples, the three-category alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled samples are calculated according to the source domain features, the prediction results of the source domain labeled samples, the first target domain features, the prediction results of the target domain labeled samples, the second target domain features and the prediction results of the target domain unlabeled samples.

[0107] It should be noted that source domain features refer to feature representations extracted from source domain labeled samples. These features can capture the intrinsic structure and key information of source domain data. By training the feature extractor, the source domain data can be mapped to a high-dimensional feature space, making the category information of the source domain data more obvious. The prediction results of source domain labeled samples refer to the results obtained by predicting source domain labeled samples after training the feature extractor and classifier on the source domain data. These prediction results reflect the classification performance of the model on the source domain data and are important indicators for evaluating the performance of the model on known data. By comparing the true labels and prediction results of the source domain labeled samples, the cross entropy classification loss of the source domain labeled samples can be calculated, thereby guiding the optimization of model parameters.

[0108] It can be understood that the first target domain feature refers to the feature representation extracted from the target domain label sample, which can also capture the intrinsic structure and key information of the target domain data. Unlike the source domain features, the target domain features may differ in distribution, so it is necessary to adapt to this difference through the training of feature extractors and classifiers. The extraction and classification recognition of target domain features are crucial for the generalization ability of the model on unseen data. The prediction results of the target domain label samples refer to the results obtained by predicting the target domain label samples after training the feature extractor and classifier on the target domain data. These prediction results reflect the classification performance of the model on the target domain data and are important indicators for evaluating the performance of the model on unseen data. By comparing the true labels and prediction results of the target domain label samples, the cross entropy classification loss of the target domain label samples can be calculated, thereby guiding the optimization of model parameters.

[0109] It is worth noting that the second target domain features refer to the feature representations extracted from the unlabeled samples in the target domain. These features can capture the potential structure and key information of the target domain data. Since these samples have no labels, the cross entropy classification loss cannot be directly calculated, but through the training of feature extractors and classifiers, these features can be made closer to the features of labeled samples in the feature space, thereby improving the generalization ability of the model on unseen data. The prediction results of the unlabeled samples in the target domain refer to the results obtained after the unlabeled samples in the target domain are classified and identified by the feature extractor and classifier. Since these samples have no real labels, the cross entropy classification loss cannot be directly calculated, but the entropy regularization loss can be used to ensure that the prediction results have a high uncertainty, thereby improving the generalization ability of the model when facing unknown data.

[0110] It can be understood that the CQAN algorithm groups the data of the two domains into four pairs of data. In adversarial learning, the discriminator learns the similarity information of each pair of inputs by judging whether each data pair comes from the same domain. In this way, the network considers the information of the source domain and the target domain at the same time, guiding the generator G s and G t In the process of adversarial learning, cross-domain domain-invariant features are learned. When the discriminator discriminates these features, it cannot correctly distinguish whether they are from the same domain. This means that the feature extractor has achieved the purpose of "deceiving" the discriminator, that is, it has achieved feature alignment of cross-domain data and can achieve similar distributions between the source domain and the target domain because the combined information from the two domains is used to narrow the distribution differences between the domains.

[0111] In a feasible implementation manner, the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category center alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the prediction probability of the target domain unlabeled sample are calculated according to the source domain features, the prediction results of the source domain label samples, the first target domain features, the prediction results of the target domain label samples, the second target domain features, and the prediction results of the target domain unlabeled samples, including: calculating according to the source domain features, the prediction results of the source domain label samples, the first target domain features, and the prediction results of the target domain label samples to obtain the cross entropy classification loss of the source domain label samples and the cross entropy classification loss of the target domain label samples; calculating according to the number of source domain label samples, the target domain label samples The method comprises the following steps: determining the source domain data class center, the target domain data class center and the combined data class center based on the number of source domain features, the first target domain features, the second target domain features and the prediction results of the target domain unlabeled samples; determining the three-class class center alignment loss based on the source domain data class center, the target domain data class center and the combined data class center; constructing a four-data pair based on the source domain features, the source domain data class center, the first target domain features and the target domain data class center; judging the four-data pair by the discriminator to generate a domain prediction result; determining the class-level four-tuple adversarial loss based on the domain prediction result and the four-data pair; and determining the entropy regularization loss of the target domain unlabeled sample prediction probability based on the number of target domain unlabeled samples and the prediction results of the target domain unlabeled samples.

[0112] It should be noted that in the CQAN algorithm, labeled source domain samples and a small number of labeled samples in the target domain are used to train the feature extractor G s and G t And the classifier G C . Due to G s (X s ) and G t (X t ) is the domain invariant feature extracted from the source domain and the target domain. In this case, the source domain with label G s (X s ) and a small number of labeled G in the target domain t (X tl ) and train the classifier G C Applying it to the unlabeled data in the target domain will have a relatively ideal effect. The classification loss of the labeled samples in the source domain and a small number of labeled samples in the target domain will help to better extract discriminative features. In semi-supervised multi-classification tasks, the cross entropy loss function is generally used to optimize the parameters of the neural network. The cross entropy loss function of a labeled sample x is as follows:

[0113]

[0114] Where y is the hot encoding vector of the label information of instance x, the dimension of y is 1×C, and only the value of the corresponding category in the hot encoding vector is 1, and the rest are 0. p is the predicted probability output of classifier C, p c The probability that instance x belongs to the cth class after passing through the entire network. f (x) represents the features extracted by the feature extractor, i.e., G f ∈{G s ,G t}. G c (G f (x) c Represents the value of the cth node of the classification network output. The S(·) function represents the softmax function, which means that the output vector will be normalized in a probabilistic form.

[0115] Labeled data X of two domains s and X tl The classification loss is as follows:

[0116]

[0117] Among them, L s is the cross entropy classification loss of the source domain label sample, L tl is the cross entropy classification loss of a small amount of labeled data in the target domain, X s is the source domain label sample, X tl is the target domain label sample, Υ s is the label of the source domain label sample, Υ tl is the label of the target domain label sample, θ s and θ t are the parameters of the two feature extractors, θ c is the classifier G C Parameters, n s is the number of source domain label samples, n tl is the number of labeled sample data in the target domain, x s i is the source domain label sample X s The i-th instance in x tl i is the target domain label sample X tl The i-th instance in y s i For label s The i-th instance in y tl i For label tl The i-th instance in G C () is the classifier G C The output of G s () and Gt () is the classifier G s and G t Extracted features.

[0118] It is understandable that there are C types of data in total, and there are C discriminators in the network, namely D c , c = 1,…,C. For each discriminator D c For example, its input is a four-dimensional data pair in represents the features of the i-th data of the c-th category in the source domain after feature extraction, represents the probability weighted feature of the j-th data of the c-th category in the target domain after feature extraction, It represents the center of the cth class in the source domain after feature extraction. Indicates the c-th class center of the target domain after feature extraction. In the process of four-element adversarial learning, all source domain data and target domain data will participate in the formation of four-element data pairs, where Represents two forms of data pairs from foreign domains; Represents two data pairs from the same domain. Each data pair in the CQAN network consists of an instance sample and a class center. This is because the discriminator in the CQAN network is more inclined to discover similar information between two data, and the similarity between the class center and each instance is easier to obtain and more robust. Because in class-level adversarial, the category information of the instance is applied. When the data pair is composed of "instance-class center", it is possible that the wrong prediction information only has a negative impact on the instance, and has little impact on the class center. If the data pair is composed of two instances, the worst case is that the prediction information of both instances may be incorrect, which has an uncontrollable impact on network training. Source domain c-th class sample Probability weighted samples of class c in the target domain Source domain class c Target domain c class heart There are only four cases where these four data form a data pair of “instance-class center”, namely, the four-data pair in the CQAN network. For each class-level discriminator D c For example, its task each time is to judge The source of these data pairs is the same. The task of the feature extractor is to make the feature differences of the quaternary data pairs extracted by it not obvious enough, resulting in D c It is impossible to correctly distinguish and achieve the purpose of cross-domain data alignment.

[0119] Each discriminator is only responsible for processing data of a certain category in the two domains. For example, the cth discriminator D c The input is the source domain and target domain prediction probability weighted feature Gs (X s )and in, Represents the predicted probability that the target domain unlabeled sample belongs to the cth class after passing through the entire network. By weighting the features, it is possible to adaptively determine the probability of each sample for the cth discriminator D c It can alleviate the negative impact of using hard pseudo labels to forcibly assign a sample to a possibly incorrect category and reduce negative transfer. The class-level four-element domain classification loss in CQAN is as follows:

[0120]

[0121] Among them, D s is the label of the same domain data pair, D t is the label of data pairs in different domains, θ s and θ t are the parameters n of the two feature extractors s is the number of source domain samples, n t is the number of samples in the target domain, D c () is the discriminator D c The output of i It is the domain label of the feature pair, indicating whether the feature pair comes from the same domain. When the feature pair is extracted from the same domain, its domain label is set to 1, otherwise the domain label is 0. f1 (x i ) and G f2 (x j ) represents the features extracted by the feature extractor, and is a unified representation of quaternary data pairs, and For instance x i and x j The predicted probabilities of belonging to the cth class, For instance x i The probability weighted features of the source domain data X s and a small amount of data X in the target domain tl They are labeled, and their probability of belonging to a certain category is either 1 or 0, so only the target domain has unlabeled data X tu Using probability weighted features, which is consistent with the four-dimensional data pair, is the domain prediction output of the c-th discriminator. If Larger, meaning that instance x i The possibility of belonging to category c is also greater. D For a feature pair (x i ,x j ) domain binary classification loss, Represents the prediction result of the c-th discriminator.

[0122] Class-level quaternary adversarial domain adaptation weights features and uses various independent discriminators to achieve adversarial alignment of various features. According to the minimax game, the learning objective of the class-level domain adversarial loss is as follows:

[0123]

[0124] In the above formula, only the influence of domain adversarial loss on parameters is considered. In fact, the parameter learning of feature extractor will also be affected by source domain cross entropy loss, target domain cross entropy loss, three-category alignment loss and target domain entropy regularization loss. In the training process of adversarial network, the discriminator hopes to distinguish whether the feature pairs come from the same data domain as much as possible, so the domain classification loss should be minimized to achieve the goal; the class-level four-element adversarial domain adaptation proposed in this chapter has C discriminators that do not affect each other. What each discriminator has to do is to distinguish the source of the cross-domain data pairs of the corresponding category, which is equivalent to C binary classifiers working at the same time. For the feature extractor, it has to maximize the loss of domain adversarial because its purpose is to deceive the discriminator. Therefore, through this minimax game, for each category, the discriminator cannot accurately distinguish the source of the feature pairs extracted by the feature extractor, and the purpose of aligning various features of the algorithm is achieved, that is, the cross-domain transfer of knowledge is realized.

[0125] Neural network learning uses the standard back-propagation formula to derive and update parameters. Most learning tasks eventually obtain a local optimal solution; however, in adversarial learning, the two optimization goals are diametrically opposed, and the final learning is very likely to be a saddle point. Therefore, for the discriminator to minimize the domain classification loss and the feature extractor to maximize the domain classification loss, CQAN uses the Gradient Reversal Layer (GRL) and places it between the feature extractor and the discriminator. GRL is an identity mapping function in forward propagation, and it reverses the gradient during back propagation. It can enable the discriminator to minimize the domain adversarial loss while maximizing the domain classification loss for the feature extractor, thus avoiding the problem of CQAN having to alternately train the feature extractor and the discriminator.

[0126] It is worth noting that in CQAN, instance samples and the adversarial class centroids of the two domains are used to learn the similarity measurement of cross-domain data pairs and thus realize cross-domain data feature alignment. Although a small amount of target domain data is labeled, it is not enough to learn a very effective knowledge model. The prediction results of the unlabeled data in the target domain in the adversarial between various instances and class centroids of the CQAN algorithm are relatively sensitive, and a small amount of misclassification has little effect on the class centroids. Therefore, a three-class centroid alignment strategy is introduced in CQAN to further improve the knowledge transfer effect. The three-class centroid alignment in CQAN is the class centroids of the source domain, the target domain, and the combined data of the source domain and the target domain. The feature distribution alignment task of cross-domain data is achieved by forcibly constraining these three class centroids. The three-class centroid alignment loss function is as follows:

[0127]

[0128] Among them, u s (c) is the class center of the c-th class data in the source domain, u t (c) is the class center of the c-th class data in the target domain, u st (c) is the centroid of the cth class of data in the combined data of the source domain and the target domain.

[0129] Calculate the class center of a data domain, that is, the average vector of all data in each category. Since a large amount of target domain data does not have real labels, calculating the class center of each category based on the predicted label will inevitably have a negative impact, because the sample has a probability of being assigned to a category that may be incorrect. The strategy adopted is to calculate the probabilistic class center. The greater the probability that an instance belongs to class c, the more likely it is to belong to this category. The greater the weight of this instance when calculating the class center of class c. Conversely, the smaller the probability, the greater the probability of incorrectness, and the smaller its contribution to calculating this class center. Therefore, the class center calculation method for the three groups of data, source domain, target domain labeled data, and target domain unlabeled data, is as follows:

[0130]

[0131]

[0132]

[0133] Among them, u s (c) is the class center of the c-th class data in the source domain, u tl (c) is the class center of the c-th class label data in the target domain, u tu (c) is the class center of the cth class of unlabeled data in the target domain, represents the number of samples of the cth category in the source domain, Indicates the number of samples of category c in the target domain labeled data, It represents the probability that the j-th unlabeled data in the target domain is classified into the c-th category by the classifier.

[0134] By minimizing the three-class center alignment loss, the source domain and target domain features can become more and more similar at the class level. In the CQAN algorithm, the domain-invariant features extracted from the two domains need to be jointly optimized by various center-class constraints and class-level four-element adversarial networks. On the one hand, the algorithm uses the prediction probability of the target domain to weight the features in order to achieve adversarial learning, so the prediction probability of the target domain greatly affects the performance of transfer learning. The use of center-class constraints can enhance the classification performance of the network and obtain more accurate prediction results, thereby improving the classification effect. On the other hand, the algorithm draws on the adversarial learning strategy, which is often not very strong in learning ability when the data distribution is very different. Especially in heterogeneous transfer learning problems, the two domain data come from different feature spaces, and the learning difficulty will be greater. By combining the center-class alignment loss to assist the class-level four-element adversarial transfer learning, better transfer effects can be achieved.

[0135] It is worth noting that the three-category alignment loss and the class-level four-element adversarial loss are both intended to minimize the distribution differences of data from different feature spaces. However, since there are few labeled data in the target domain, the distribution estimate of each type of data in the target domain may be incorrect. Therefore, the output probability of the target domain is introduced as an entropy regularization loss, because the algorithm expects the prediction probability of the network to present a peak distribution, and does not want it to obey a flat distribution, that is, the classification result should clearly belong to a certain category without fuzzy classification. The entropy regularization loss of the prediction probability of unlabeled samples in the target domain is as follows:

[0136]

[0137] Among them, X tu is the unlabeled sample in the target domain, n tu is the number of unlabeled samples in the target domain, Represents the unlabeled i-th data in the target domain The probability of belonging to class c.

[0138] The entropy regularization loss of the prediction probability of unlabeled samples in the target domain only constrains the unlabeled samples in the target domain. This is because the prediction probability loss of all samples in the source domain and a small number of labeled samples in the target domain will not have a flat distribution, because they are truly labeled and their prediction probabilities are constrained by the supervised cross entropy loss.

[0139] In the specific implementation, the CQAN algorithm uses a deep network structure and a back-propagation chain derivation method to update the parameters of the feature extractor, classifier, and discriminator to achieve the best performance and accuracy. The implementation steps of CQAN include: s and label Y s , a small amount of labeled data X in the target domain tl and label Y tl , unlabeled data X in the target domain tu , learning rate η, balance parameters λ, α, β, maximum number of iterations M as input, and randomly initialized parameters θ s ,θ t ,θ c , Train the feature extractor G s , G t , classifier G C and the discriminator Use the trained feature extractor G s , G t and classifier G C For the target domain data X tu Perform classification and recognition, and output the target domain data X tu Tags The CQAN framework reduces the need for large-scale annotated datasets through transfer learning, which is achieved by using existing data for adversarial training to transfer knowledge to new environmental perception tasks, thereby reducing the cost of data collection and annotation. Independently trained class-level discriminators perform fine-grained classification and alignment of features of different categories, which helps capture the feature differences between different categories and improves the model's recognition accuracy of environmental elements.

[0140] In a feasible implementation manner, the determining of feature extractor parameters, discriminator parameters and classifier parameters according to at least one of the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the target domain unlabeled sample prediction probability includes: determining feature extractor parameters according to the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the target domain unlabeled sample prediction probability; determining discriminator parameters according to the three-category center alignment loss; determining classifier parameters according to the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample and the entropy regularization loss of the target domain unlabeled sample prediction probability.

[0141] It should be noted that the discriminator D cThe task is to discriminate whether the data pairs input each time are from the same domain or different domains. The CQAN network pays more attention to whether the data pairs are from the same domain, rather than whether the data pairs from the same domain are specifically from the source domain or the target domain. In this way, the network will be more inclined to explore the similarities between the data. The feature extractor uses the similarity measure learned in the network to confuse the judgment of the discriminator, making it impossible for the discriminator to accurately distinguish between data pairs from the same domain and different domains, so as to achieve the purpose of cross-domain data feature alignment. However, in some classification tasks, such as hyperspectral remote sensing classification tasks, the differences between different types of data are large, and the overall adversarial strategy cannot guarantee that the migration effect of each type of data is ideal. Therefore, considering that each class must undergo four-element adversarial learning, CQAN uses C discriminators to achieve class-level four-element adversarial domain adaptation. The total loss function of CQAN is expressed as follows:

[0142]

[0143] Among them, the first term L s is the cross entropy classification loss of the source domain label sample, and the second term L tl is the cross entropy classification loss of a small amount of labeled data in the target domain, and the third term L cla_mean is the three-category alignment loss function, and the fourth term L D is the class-level quaternary adversarial loss, and the last term L entropy is the entropy regularization loss of the target domain data prediction probability, λ1, λ2, α, β are hyperparameters, θ s and θ t are the parameters of the two feature extractors, θ c is the classifier G C Parameters, is the parameter of the c-th discriminator.

[0144] In the CQAN network, the parameters θ of the two feature extractors are s and θ t It is constrained by five losses, and the classifier G C The parameters of are jointly determined by the cross entropy loss of the source domain and the target domain and the entropy regularization loss of the target domain prediction probability, and the parameters of the discriminator are only learned by the class-level four-tuple domain adversarial loss. The goal of the total loss is to find the local optimal solution and As follows:

[0145]

[0146] In this embodiment, the source domain data and the target domain data are obtained by dividing and labeling the historical surrounding environment data after preprocessing, and then adversarial training and transfer learning are performed on the class-level four-element adversarial network model based on the source domain data and the target domain data, thereby achieving effective alignment of data feature differences under different driving environments and effectively improving the generalization ability and adaptability of the model.

[0147] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the vehicle environment perception method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0148] This application also provides a vehicle environment perception device, please refer to Figure 5 , the vehicle environment perception device comprises:

[0149] The optimization module 10 is used to obtain historical surrounding environment data of the vehicle, and optimize the class-level four-element adversarial network model according to the historical surrounding environment data to obtain an environmental perception model.

[0150] The prediction module 20 is used to perform classification and identification based on the environmental perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target.

[0151] The perception module 30 is used to perform environmental perception according to the predicted probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information.

[0152] The execution module 40 is used to execute corresponding driving decisions according to the environmental perception information.

[0153] The present embodiment provides a vehicle environment perception device. The present embodiment optimizes a class-level four-element adversarial network model according to the historical surrounding environment data of the vehicle to obtain an environment perception model; performs classification and identification according to the environment perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target; performs environment perception according to the prediction probability corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environment perception information; and executes corresponding driving decisions according to the environment perception information, which can effectively improve the perception ability of the car under different environmental conditions and enhance the perception accuracy.

[0154] In summary, this embodiment optimizes the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle, thereby effectively improving the adaptability and recognition ability of the model to environmental changes. The predicted probability corresponding to each category of the detection target is quickly and accurately determined through the environmental perception model and the real-time surrounding environment data of the vehicle, and then the environment is perceived according to the predicted probability and the current natural environment conditions of the vehicle. The perceived environment perception information can be accurately obtained and the corresponding driving decisions can be executed, which overcomes the technical defects of the environmental perception system of the autonomous driving vehicle in quickly adapting to changing environmental conditions, and can effectively improve the perception ability of the vehicle under different environmental conditions and enhance the perception accuracy.

[0155] Optionally, the optimization module 10 is also used to obtain historical surrounding environment data of the vehicle, and perform preprocessing operations on the historical surrounding environment data to obtain preprocessed historical surrounding environment data, wherein the preprocessing operations at least include data cleaning, data standardization and noise removal operations; divide and label the preprocessed historical surrounding environment data to obtain source domain data and target domain data, wherein the source domain data is composed of source domain labeled samples, and the target domain data is composed of target domain labeled samples and target domain unlabeled samples; optimize the class-level four-element adversarial network model according to the source domain data and the target domain data to obtain an environmental perception model, wherein the class-level four-element adversarial network model includes a feature extractor, a discriminator and a classifier.

[0156] Optionally, the optimization module 10 is also used to train the class-level four-element adversarial network model according to the source domain data and the target domain data to obtain model performance parameters; and optimize the adversarial network model according to the model performance parameters to obtain an environmental perception model.

[0157] Optionally, the optimization module 10 is further used to train the feature extractor, the discriminator and the classifier according to the source domain data and the target domain data, respectively, to obtain the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the target domain unlabeled sample prediction probability; determine the feature extractor parameters, the discriminator parameters and the classifier parameters according to at least one of the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the target domain unlabeled sample prediction probability; determine the model performance parameters according to the feature extractor parameters, the discriminator parameters and the classifier parameters.

[0158] Optionally, the optimization module 10 is further used to train the feature extractor and the classifier through the source domain label samples and the target domain label samples to obtain source domain features, prediction results of source domain label samples, first target domain features and prediction results of target domain label samples; perform feature extraction and classification identification on the target domain unlabeled samples through the trained feature extractor and the trained classifier to obtain second target domain features and prediction results of target domain unlabeled samples; calculate the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled samples according to the source domain features, the prediction results of the source domain label samples, the first target domain features, the prediction results of the target domain label samples, the second target domain features and the prediction results of the target domain unlabeled samples.

[0159] Optionally, the optimization module 10 is further used to calculate according to the source domain features, the prediction results of the source domain label samples, the first target domain features and the prediction results of the target domain label samples to obtain the cross entropy classification loss of the source domain label samples and the cross entropy classification loss of the target domain label samples; determine the source domain data class center, the target domain data class center and the combined data class center according to the number of source domain label samples, the number of target domain label samples, the source domain features, the first target domain features, the second target domain features and the prediction results of the target domain unlabeled samples; determine the three-class center alignment loss according to the source domain data class center, the target domain data class center and the combined data class center; construct a four-dimensional data pair according to the source domain features, the source domain data class center, the first target domain features and the target domain data class center; judge the four-dimensional data pair through the discriminator to generate a domain prediction result; determine the class-level four-dimensional adversarial loss according to the domain prediction result and the four-dimensional data pair; determine the entropy regularization loss of the target domain unlabeled sample prediction probability according to the number of target domain unlabeled samples and the prediction results of the target domain unlabeled samples.

[0160] Optionally, the optimization module 10 is further used to determine feature extractor parameters according to the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, the three-category heart alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the target domain unlabeled sample prediction probability; determine discriminator parameters according to the three-category heart alignment loss; determine classifier parameters according to the cross entropy classification loss of the source domain label samples, the cross entropy classification loss of the target domain label samples, and the entropy regularization loss of the target domain unlabeled sample prediction probability.

[0161] The vehicle environment perception device provided by the present application adopts the vehicle environment perception method in the above embodiment, which can solve the technical problem of how to effectively improve the perception ability of the car under different environmental conditions and enhance the perception accuracy. Compared with the prior art, the beneficial effects of the vehicle environment perception device provided by the present application are the same as the beneficial effects of the vehicle environment perception method provided by the above embodiment, and the other technical features in the vehicle environment perception device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0162] The present application provides a vehicle environment perception device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the vehicle environment perception method in the above-mentioned embodiment one.

[0163] Reference below Figure 6 , which shows a schematic diagram of the structure of a vehicle environment perception device suitable for implementing an embodiment of the present application. The vehicle environment perception device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The vehicle environment perception device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0164] like Figure 6As shown, the vehicle environment perception device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the vehicle environment perception device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the vehicle environment perception device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows a vehicle environment perception device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.

[0165] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0166] The vehicle environment perception device provided by the present application adopts the vehicle environment perception method in the above embodiment, which can solve the technical problem of how to effectively improve the perception ability of the car under different environmental conditions and enhance the perception accuracy. Compared with the prior art, the beneficial effects of the vehicle environment perception device provided by the present application are the same as the beneficial effects of the vehicle environment perception method provided by the above embodiment, and the other technical features in the vehicle environment perception device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0167] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0168] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0169] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the vehicle environment perception method in the above-mentioned embodiment.

[0170] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0171] The computer-readable storage medium may be included in the vehicle environment sensing device; or may exist independently without being assembled into the vehicle environment sensing device.

[0172] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the vehicle environment perception device, the vehicle environment perception device: optimizes the class-level four-element adversarial network model according to the historical surrounding environment data of the vehicle to obtain an environment perception model; performs classification and identification according to the environment perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target; performs environmental perception according to the prediction probability corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information; and executes corresponding driving decisions according to the environmental perception information.

[0173] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0174] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0175] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0176] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned vehicle environment perception method, and can solve the technical problem of how to effectively improve the perception ability of the automobile under different environmental conditions and enhance the perception accuracy. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the vehicle environment perception method provided by the above-mentioned embodiment, and will not be repeated here.

[0177] The present application also provides a computer program product, including a computer program, which implements the steps of the vehicle environment perception method as described above when executed by a processor.

[0178] The computer program product provided by this application can solve the technical problem of how to effectively improve the perception ability of the car under different environmental conditions and enhance the perception accuracy. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as the beneficial effects of the vehicle environment perception method provided by the above embodiment, which will not be repeated here.

[0179] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A vehicle environment perception method, characterized in that: The method comprises: According to the historical surrounding environment data of the vehicle, the class-level quaternary adversarial network model is optimized to obtain the environment perception model; Classify and identify the environment according to the environment perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target; Performing environmental perception according to the predicted probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information; Execute corresponding driving decisions according to the environmental perception information.

2. The method according to claim 1, characterized in that The method of optimizing the class-level quaternary adversarial network model according to the historical surrounding environment data of the vehicle to obtain the environment perception model includes: Acquire historical surrounding environment data of the vehicle, and perform a preprocessing operation on the historical surrounding environment data to obtain preprocessed historical surrounding environment data, wherein the preprocessing operation at least includes data cleaning, data standardization, and noise removal operations; Dividing and labeling the preprocessed historical surrounding environment data to obtain source domain data and target domain data, wherein the source domain data is composed of source domain labeled samples, and the target domain data is composed of target domain labeled samples and target domain unlabeled samples; The class-level four-element adversarial network model is optimized according to the source domain data and the target domain data to obtain an environmental perception model, wherein the class-level four-element adversarial network model includes a feature extractor, a discriminator, and a classifier.

3. The method according to claim 2, characterized in that The optimizing the class-level four-element adversarial network model according to the source domain data and the target domain data to obtain an environment perception model includes: Training the class-level quaternary adversarial network model according to the source domain data and the target domain data to obtain model performance parameters; The adversarial network model is optimized according to the model performance parameters to obtain an environmental perception model.

4. The method according to claim 3, characterized in that The step of training the class-level quaternary adversarial network model according to the source domain data and the target domain data to obtain model performance parameters includes: The feature extractor, the discriminator and the classifier are trained according to the source domain data and the target domain data respectively to obtain the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category heart alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled sample; Determine a feature extractor parameter, a discriminator parameter, and a classifier parameter according to at least one of the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-class center alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the prediction probability of the target domain unlabeled sample; Model performance parameters are determined based on the feature extractor parameters, the discriminator parameters, and the classifier parameters.

5. The method according to claim 4, characterized in that The feature extractor, the discriminator and the classifier are trained respectively according to the source domain data and the target domain data to obtain the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category heart alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled sample, including: The feature extractor and the classifier are trained by using the source domain label samples and the target domain label samples to obtain source domain features, prediction results of source domain label samples, first target domain features, and prediction results of target domain label samples; Performing feature extraction and classification recognition on the target domain unlabeled samples by using the trained feature extractor and the trained classifier to obtain a second target domain feature and a prediction result of the target domain unlabeled samples; According to the source domain features, the prediction results of the source domain labeled samples, the first target domain features, the prediction results of the target domain labeled samples, the second target domain features and the prediction results of the target domain unlabeled samples, the cross entropy classification loss of the source domain labeled samples, the cross entropy classification loss of the target domain labeled samples, the three-category center alignment loss, the class-level four-element adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled samples are calculated.

6. The method according to claim 5, characterized in that The method of calculating the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the prediction probability of the target domain unlabeled sample according to the source domain feature, the prediction result of the source domain label sample, the first target domain feature, the prediction result of the target domain label sample, the second target domain feature, and the prediction result of the target domain unlabeled sample includes: Calculating according to the source domain features, the prediction results of the source domain label samples, the first target domain features, and the prediction results of the target domain label samples to obtain the cross entropy classification loss of the source domain label samples and the cross entropy classification loss of the target domain label samples; Determine the source domain data class center, the target domain data class center, and the combined data class center according to the number of source domain labeled samples, the number of target domain labeled samples, the source domain features, the first target domain features, the second target domain features, and the prediction results of the target domain unlabeled samples; Determine the three-category center alignment loss according to the source domain data center, the target domain data center, and the combined data center; Constructing a four-dimensional data pair according to the source domain feature, the source domain data class center, the first target domain feature, and the target domain data class center; The discriminator judges the four-dimensional data pair to generate a domain prediction result; Determining a class-level quadruple adversarial loss based on the domain prediction result and the quadruple data pair; An entropy regularization loss of the prediction probability of the target domain unlabeled samples is determined according to the number of the target domain unlabeled samples and the prediction results of the target domain unlabeled samples.

7. The method according to claim 4, characterized in that The determining of the feature extractor parameters, the discriminator parameters and the classifier parameters according to at least one of the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-category center alignment loss, the class-level quaternary adversarial loss and the entropy regularization loss of the prediction probability of the target domain unlabeled sample includes: Determine a feature extractor parameter according to the cross entropy classification loss of the source domain label sample, the cross entropy classification loss of the target domain label sample, the three-class center alignment loss, the class-level four-element adversarial loss, and the entropy regularization loss of the target domain unlabeled sample prediction probability; Determine the discriminator parameters according to the three types of heart alignment losses; The classifier parameters are determined according to the cross entropy classification loss of the source domain labeled samples, the cross entropy classification loss of the target domain labeled samples, and the entropy regularization loss of the predicted probability of the target domain unlabeled samples.

8. A vehicle environment perception device, characterized in that: The vehicle environment perception device comprises: An optimization module, used to obtain historical surrounding environment data of the vehicle, and optimize the class-level quaternary adversarial network model according to the historical surrounding environment data to obtain an environment perception model; A prediction module is used to perform classification and identification based on the environmental perception model and the real-time surrounding environment data of the vehicle to obtain the prediction probability corresponding to each category of the detection target; A perception module, used to perform environmental perception according to the predicted probabilities corresponding to each category of the detection target and the current natural environment conditions of the vehicle to obtain environmental perception information; An execution module is used to execute corresponding driving decisions according to the environmental perception information.

9. A vehicle environment perception device, characterized in that: The vehicle environment perception device includes: a memory, a processor, and a vehicle environment perception program stored in the memory and executable on the processor, wherein the vehicle environment perception program is configured to implement the vehicle environment perception method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a vehicle environment perception program, and when the vehicle environment perception program is executed by the processor, the vehicle environment perception method according to any one of claims 1 to 7 is implemented.

Images