An intelligent control system and method for LED car lights based on environmental perception

Abstract

The present invention relates to the field of vehicle light control, and in particular to an intelligent control system and method for LED vehicle lights based on environmental perception. An intelligent control system for LED vehicle lights based on environmental perception comprises: an environmental state data acquisition module, a vehicle light control strategy output module, a vehicle light control module, a vehicle light control strategy output model real-time training module, and a value assessment network model real-time training module. The present invention outputs the LED vehicle light control strategy in different environments through a vehicle light control strategy output model, and can realize automatic control of LED vehicle lights in different environments without manual control by the driver; and with the assistance of a target vehicle light control strategy output model, the vehicle light control strategy output model is trained in real time, which can effectively improve the accuracy of the LED vehicle light control strategy and enhance the user experience of intelligent control of LED vehicle lights.

Description

Technical Field

[0001] The present invention relates to the field of vehicle light control, and in particular to an intelligent control system and method for LED vehicle lights based on environment perception. Background Art

[0002] In the modern automotive industry, the development of LED headlight technology has evolved from traditional lighting functions to a more intelligent and multifunctional trend. LED headlights not only provide more efficient and clearer lighting effects, but also have the ability to dynamically adjust brightness, color temperature and mode. This intelligent LED headlight can not only enhance the driver's visibility and driving safety, but also provide a more suitable lighting environment under different driving conditions, thereby improving the driving experience.

[0003] However, traditional LED lighting systems usually lack the ability to perceive the surrounding environment and driving conditions, resulting in the inability to make optimal lighting adjustments in complex driving scenarios. The driver still needs to make manual adjustments, affecting the driving experience. Summary of the invention

[0004] The present invention outputs LED headlight control strategies under different environments through a headlight control strategy output model, and can realize automatic control of LED headlights under different environments without the need for manual control by the driver; and with the assistance of the target headlight control strategy output model, the headlight control strategy output model is trained in real time, which can effectively improve the accuracy of the LED headlight control strategy and enhance the user experience of the LED headlight intelligent control.

[0005] An intelligent control method for LED vehicle lights based on environmental perception, comprising:

[0006] Acquire environmental status data during vehicle driving at the monitoring time point;

[0007] Send the current environment status data to the headlight control strategy output model for processing, and output the LED headlight control strategy;

[0008] Intelligently control the LED lights on the vehicle through LED light control strategies;

[0009] At the same time, it also includes real-time training of the output model of the headlight control strategy. The specific steps are as follows:

[0010] At the monitoring time point, the environmental state data is sent to the target headlight control strategy output model for processing, and the target headlight control strategy is output. Then, the environmental state data and the target headlight control strategy are spliced ​​to construct value assessment data, and the value assessment data is sent to the value assessment network model for processing, and the strategy value data is output. The strategy gradient value of the strategy value data for the target headlight control strategy is calculated, and then the parameters of the headlight control strategy output model are adjusted using the gradient ascent method based on the strategy gradient value, so as to realize the real-time training of the headlight control strategy output model; at the same time, it also includes the real-time training of the value assessment network model;

[0011] At the target network update time point, the headlight control strategy output model is directly used as the target headlight control strategy output model to replace the previous target headlight control strategy output model, and N monitoring time points are included between two adjacent target network update time points, and the time period between two adjacent target network update time points is recorded as the analysis period.

[0012] As a preferred aspect of the present invention, real-time training is performed on the value assessment network model, specifically comprising the following steps:

[0013] Obtain all value assessment data acquired in the current analysis cycle to form a local buffer value assessment database, and at the same time obtain the current geographic location of the vehicle, and match the cloud buffer value assessment database in the current analysis cycle from the cloud server based on the current geographic location of the vehicle;

[0014] All the value assessment data in the local buffer value assessment database and the cloud buffer value assessment database within the current analysis cycle are sequentially input into the target value assessment network model for processing to obtain several target strategy value data, and the largest target strategy value data among all the target strategy value data is selected as the identification strategy value data. The loss value is calculated based on the identification strategy value data and the strategy value data, and the parameters of the value assessment network model are adjusted through the back propagation algorithm based on the loss value to realize real-time training of the value assessment network model.

[0015] As a preferred aspect of the present invention, matching a cloud buffer value assessment database within the current analysis period from a cloud server based on the current geographical location of the vehicle specifically includes the following steps:

[0016] For any vehicle, all value assessment data acquired in the current analysis cycle are annotated by geographic location, cloud buffer value assessment data is constructed and uploaded to the cloud server;

[0017] Calculate the distance between the vehicle's current geographic location and the geographic location corresponding to all cloud buffer value assessment data, record it as the regional distance, compare all regional distances with the preset distance, and form a cloud buffer value assessment database with the cloud buffer value assessment data corresponding to all regional distances less than the preset distance to complete the matching of the cloud buffer value assessment database.

[0018] As a preferred aspect of the present invention, training is performed on the vehicle light control strategy output model, specifically comprising the following steps:

[0019] A plurality of headlight control strategy output training samples are obtained, the headlight control strategy output training samples include environmental state data and its corresponding optimal headlight control strategy, all headlight control strategy output training samples are combined into a headlight control strategy output training set, and the headlight control strategy output training set is sent to a headlight control strategy output model with initialized parameters for training. During the training, the environmental state data is used as input, and the optimal headlight control strategy is used as a target. The first accuracy rate corresponding to the headlight control strategy output model is calculated, and it is determined whether the first accuracy rate corresponding to the headlight control strategy output model is higher than the accuracy threshold. If the first accuracy rate corresponding to the headlight control strategy output model is higher than the accuracy threshold, the trained headlight control strategy output model is output, and the initial headlight control strategy output model and the target headlight control strategy output model are both trained headlight control strategy output models; otherwise, the headlight control strategy output model is continuously trained through the headlight control strategy output training set.

[0020] As a preferred aspect of the present invention, the training of the value assessment network model specifically includes the following steps:

[0021] A number of positive value assessment training samples and negative value assessment training samples are obtained, wherein the positive value assessment training samples include environmental state data and its corresponding optimal headlight control strategy, and the negative value assessment training samples include environmental state data and its corresponding worst headlight control strategy. The positive value assessment training samples and the negative value assessment training samples are marked by the target value, and all the marked positive value assessment training samples and the negative value assessment training samples form a value assessment training set, and the value assessment training set is sent to the parameter-initialized value assessment network model for training. During the training, the positive value assessment training samples or the negative value assessment training samples are used as input, and the target value is used as the target. The second accuracy rate corresponding to the value assessment network model is calculated, and it is determined whether the second accuracy rate corresponding to the value assessment network model is higher than the accuracy threshold. If the second accuracy rate corresponding to the value assessment network model is higher than the accuracy threshold, the trained value assessment network model is output, and the initial value assessment network model and the target value assessment network model are both trained value assessment network models; otherwise, the value assessment network model is continuously trained by the value assessment training set; and the hyperparameters of the value assessment network model are adjusted by a genetic algorithm.

[0022] As a preferred aspect of the present invention, the vehicle light control strategy output model and the value evaluation network model are both established based on the BP neural network model.

[0023] As a preferred aspect of the present invention, the hyperparameters of the vehicle light control strategy output model and the hyperparameters of the value assessment network model are adjusted by genetic algorithms.

[0024] The present invention also provides an LED vehicle light intelligent control system based on environmental perception, comprising:

[0025] An environmental status data acquisition module is used to acquire environmental status data during vehicle driving at a monitoring time point;

[0026] The headlight control strategy output module is used to send the current environment status data to the headlight control strategy output model for processing and output the LED headlight control strategy;

[0027] A headlight control module is used to intelligently control the LED headlights on the vehicle through an LED headlight control strategy;

[0028] The real-time training module of the headlight control strategy output model is used to send the environmental state data to the target headlight control strategy output model for processing at the monitoring time point, output the target headlight control strategy, then splice the environmental state data with the target headlight control strategy, construct value assessment data, send the value assessment data to the value assessment network model for processing, output the strategy value data, calculate the strategy gradient value of the strategy value data for the target headlight control strategy, and then adjust the parameters of the headlight control strategy output model using the gradient ascent method based on the strategy gradient value, so as to realize the real-time training of the headlight control strategy output model;

[0029] The value assessment network model real-time training module is used to perform real-time training on the value assessment network model.

[0030] The present invention has the following advantages:

[0031] 1. The present invention outputs the LED headlight control strategy under different environments through the headlight control strategy output model, which can realize automatic control of LED headlights under different environments without the need for manual control by the driver; and with the assistance of the target headlight control strategy output model, the headlight control strategy output model is trained in real time, which can effectively improve the accuracy of the LED headlight control strategy and enhance the user experience of LED headlight intelligent control.

[0032] 2. The present invention obtains the cloud buffer value assessment database and the local buffer value assessment database of other vehicles to perform real-time training on the value assessment network model, and can learn the headlight control strategies of different vehicles in the current area. The target headlight control strategy output model is trained in real time with the real-time trained value assessment network model, which can timely adapt to environmental changes in different regions and further improve the accuracy of the LED headlight control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the structure of an engineering safety prediction system based on the Internet of Things adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0035] Embodiment 1, a method for intelligently controlling LED vehicle lights based on environment perception, comprising:

[0036] At the monitoring time point, the environmental status data of the vehicle during driving is obtained. The environmental status data includes current visibility, weather conditions, current time, distance to the vehicle in front, headlight status of the vehicle in front and distance to the oncoming vehicle, etc. The current visibility can be obtained through a camera or a laser radar, the current weather conditions can be obtained through a meteorological sensor, the current time can be obtained through a system clock, the distance to the vehicle in front and the distance to the oncoming vehicle can be obtained through a laser radar, and the headlight status can be obtained through a camera. These environmental data can reflect the current environmental conditions of the vehicle and realize environmental perception;

[0037] The current environment state data is sent to the headlight control strategy output model for processing. The headlight control strategy output model is established based on the BP neural network and outputs the LED headlight control strategy. It should be noted that the LED headlight control strategy specifically includes various headlights and their corresponding PWM control information;

[0038] Intelligently control the LED lights on the vehicle through LED light control strategies;

[0039] At the same time, it also includes real-time training of the output model of the headlight control strategy. The specific steps are as follows:

[0040] At the monitoring time point, the environmental state data is sent to the target headlight control strategy output model for processing, and the target headlight control strategy is output. The environmental state data and the target headlight control strategy are then spliced ​​to construct value assessment data, and the value assessment data is sent to the value assessment network model for processing. The value assessment network model is established based on the BP neural network model, and the strategy value data is output. The strategy value data can reflect the value corresponding to the target headlight control strategy under the environment corresponding to the environmental state data, that is, the feasibility of using the target headlight control strategy to control the LED headlights, and the strategy gradient value of the strategy value data for the target headlight control strategy is calculated. The specific calculation method can be to calculate the derivative of the strategy value data for the target headlight control strategy, and use the calculated derivative as the strategy gradient value, and then use the gradient ascent method based on the strategy gradient value to adjust the parameters of the headlight control strategy output model to realize real-time training of the headlight control strategy output model; at the same time, it also includes real-time training of the value assessment network model;

[0041] At the target network update time point, the headlight control strategy output model is directly used as the target headlight control strategy output model to replace the previous target headlight control strategy output model, and the time between two adjacent target network update time points includes N monitoring time points, and the time period between two adjacent target network update time points is recorded as the analysis period, and the value of N is determined by the developer. It should be noted that between two adjacent target network update time points, the headlight control strategy output model will be updated based on the target headlight control strategy output model. Under the correction of the target headlight control strategy output model, it can effectively avoid the headlight control strategy output model from updating too much, which affects the accuracy of the LED headlight control strategy output;

[0042] The present application outputs the LED headlight control strategy in different environments through the headlight control strategy output model, which can realize automatic control of the LED headlights in different environments without the need for manual control by the driver; and with the assistance of the target headlight control strategy output model, the headlight control strategy output model is trained in real time, which can effectively improve the accuracy of the LED headlight control strategy and enhance the user experience of the LED headlight intelligent control.

[0043] Real-time training of the value assessment network model includes the following steps:

[0044] Obtain all value assessment data acquired in the current analysis cycle to form a local buffer value assessment database, and at the same time obtain the current geographic location of the vehicle, and match the cloud buffer value assessment database in the current analysis cycle from the cloud server based on the current geographic location of the vehicle. It should be noted that when any vehicle outputs value assessment data, a corresponding cloud buffer value assessment database will be constructed and uploaded to the cloud server. The cloud buffer value assessment database in the cloud server can be accessed by all authorized vehicles;

[0045] All the value assessment data in the local buffer value assessment database and the cloud buffer value assessment database within the current analysis cycle are sequentially input into the target value assessment network model for processing to obtain several target strategy value data, and the largest target strategy value data among all the target strategy value data is selected as the identification strategy value data. The loss value is calculated based on the identification strategy value data and the strategy value data, and the parameters of the value assessment network model are adjusted through the back propagation algorithm based on the loss value to realize real-time training of the value assessment network model.

[0046] This application obtains the cloud buffer value assessment database and local buffer value assessment database of other vehicles to perform real-time training on the value assessment network model, and can learn the headlight control strategies of different vehicles in the current area. The target headlight control strategy output model is trained in real time with the real-time trained value assessment network model, which can achieve timely adaptation to environmental changes in different regions and further improve the accuracy of LED headlight control strategy.

[0047] Based on the current geographic location of the vehicle, the cloud buffer value assessment database within the current analysis period is matched from the cloud server, specifically including the following steps:

[0048] For any vehicle, all value assessment data acquired in the current analysis cycle are annotated by geographic location, cloud buffer value assessment data is constructed and uploaded to the cloud server;

[0049] Calculate the distance between the vehicle's current geographic location and the geographic location corresponding to all cloud buffer value assessment data, recorded as the regional distance. The distance calculation method can use Euclidean distance. All regional distances are compared with the preset distance. The preset distance is determined by the developer, and the cloud buffer value assessment data corresponding to all regional distances less than the preset distance are combined into a cloud buffer value assessment database to complete the matching of the cloud buffer value assessment database.

[0050] Training the output model of the vehicle light control strategy includes the following steps:

[0051] Obtain several sets of headlight control strategy output training samples, which include environmental state data and its corresponding optimal headlight control strategy. It should be noted that the headlight control strategy output training samples are obtained by developers based on vehicle simulation driving tests. Under the vehicle simulation driving test based on simulation software, developers will record the optimal headlight control strategy under different environments. For example, when visibility is low, turn on the high beam, and form a headlight control strategy output training set with all headlight control strategy output training samples. The headlight control strategy output training set is then sent to the headlight control strategy output model with parameter initialization for training. During the training, the environmental state data is used as input. Input, take the optimal headlight control strategy as the goal, calculate the first accuracy corresponding to the headlight control strategy output model, and judge whether the first accuracy corresponding to the headlight control strategy output model is higher than the accuracy threshold. The accuracy threshold is determined by the developer. If the first accuracy corresponding to the headlight control strategy output model is higher than the accuracy threshold, output the trained headlight control strategy output model, and the initial headlight control strategy output model and the target headlight control strategy output model are both trained headlight control strategy output models; otherwise, continue to train the headlight control strategy output model through the headlight control strategy output training set; the hyperparameters of the headlight control strategy output model are adjusted through the genetic algorithm.

[0052] Here we briefly describe the steps of adjusting the hyperparameters of the headlight control strategy output model through genetic algorithm. First, we will set the simulated individuals based on the hyperparameters of the headlight control strategy output model. The hyperparameters of the headlight control strategy output model may include the number of hidden layers and the number of neural nodes in the hidden layers. For each simulated individual, the hyperparameter values ​​stored are random within the set range. Then, all simulated individuals are grouped into a population set, and the fitness value corresponding to each simulated individual is calculated. The specific method is to set the headlight control strategy output model according to the hyperparameters corresponding to the simulated individual. Then, the headlight control strategy output model with the set hyperparameters is trained, and the accuracy of the headlight control strategy output model is calculated with the headlight control strategy output test set. This accuracy is the fitness value corresponding to the simulated individual, and the form of the headlight control strategy output test set is consistent with the headlight control strategy output training set. Based on the fitness value corresponding to each simulated individual, the population set is iteratively simulated through the genetic algorithm. The specific operations include selection, recombination and mutation. When a certain number of iterations is reached, the simulated individual with the largest output fitness value sets the hyperparameters of the headlight control strategy output model.

[0053] The training of the value assessment network model includes the following steps:

[0054] Obtain several positive samples and negative samples for value assessment training. The positive samples for value assessment training include environmental state data and its corresponding optimal headlight control strategy. The negative samples for value assessment training include environmental state data and its corresponding worst headlight control strategy. It should be noted that when developers conduct vehicle simulation driving tests, they will record the optimal headlight control strategies in different environments to form positive samples for value assessment training, and record the worst headlight control strategies in different environments to form negative samples for value assessment training. The worst headlight control strategy here refers to a headlight control strategy that completely deviates from driving rules. The positive samples and negative samples for value assessment training are marked by target value. Generally, the target value of the positive samples for value assessment training is 1, and the target value of the negative samples for value assessment training is 0. All marked positive samples for value assessment training are marked as 1. The positive samples and the negative samples of value assessment training form a value assessment training set, and the value assessment training set is sent to the value assessment network model with parameter initialization for training. During the training, the positive samples of value assessment training or the negative samples of value assessment training are used as input, and the target value is used as the target. The second accuracy corresponding to the value assessment network model is calculated to determine whether the second accuracy corresponding to the value assessment network model is higher than the accuracy threshold. The accuracy threshold is determined by the developer. If the second accuracy corresponding to the value assessment network model is higher than the accuracy threshold, the trained value assessment network model is output, and the initial value assessment network model and the target value assessment network model are both trained value assessment network models; otherwise, the value assessment network model is continuously trained through the value assessment training set; the hyperparameters of the value assessment network model are adjusted through a genetic algorithm.

[0055] Embodiment 2, an intelligent control system for LED lights based on environmental perception, such as Figure 1 As shown, including:

[0056] The environment status data acquisition module is used to acquire the environment status data of the vehicle during driving at the monitoring time point. The environment status data includes current visibility, weather conditions, current time, distance to the vehicle in front, headlight status of the vehicle in front and distance to the oncoming vehicle, etc. The current visibility can be acquired through a camera or a laser radar, the current weather conditions can be acquired through a meteorological sensor, the current time can be acquired through a system clock, the distance to the vehicle in front and the distance to the oncoming vehicle can be acquired through a laser radar, and the headlight status can be acquired through a camera. These environment data can reflect the current environment conditions of the vehicle and realize environment perception;

[0057] The headlight control strategy output module is used to send the current environment state data to the headlight control strategy output model for processing. The headlight control strategy output model is established based on the BP neural network and outputs the LED headlight control strategy. It should be noted that the LED headlight control strategy specifically includes various headlights and their corresponding PWM control information;

[0058] A headlight control module is used to intelligently control the LED headlights on the vehicle through an LED headlight control strategy;

[0059] The real-time training module of the headlight control strategy output model is used to send the environmental state data to the target headlight control strategy output model for processing at the monitoring time point, output the target headlight control strategy, and then splice the environmental state data with the target headlight control strategy to construct value assessment data, and send the value assessment data to the value assessment network model for processing. The value assessment network model is established based on the BP neural network model, and outputs strategy value data. The strategy value data can reflect the value corresponding to the target headlight control strategy under the environment corresponding to the environmental state data, that is, the feasibility of using the target headlight control strategy to control the LED headlights, and calculate the strategy gradient value of the strategy value data for the target headlight control strategy. The specific calculation method can be to calculate the derivative of the strategy value data for the target headlight control strategy, and use the calculated derivative as the strategy gradient value, and then use the gradient ascent method based on the strategy gradient value to adjust the parameters of the headlight control strategy output model to realize real-time training of the headlight control strategy output model;

[0060] The value assessment network model real-time training module is used to perform real-time training on the value assessment network model.

[0061] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An intelligent control method for LED car lights based on environmental perception, characterized in that: include: Acquire environmental status data during vehicle driving at the monitoring time point; Send the current environment status data to the headlight control strategy output model for processing, and output the LED headlight control strategy; Intelligently control the LED lights on the vehicle through LED light control strategies; At the same time, it also includes real-time training of the output model of the headlight control strategy. The specific steps are as follows: At the monitoring time point, the environmental state data is sent to the target headlight control strategy output model for processing, and the target headlight control strategy is output. Then, the environmental state data and the target headlight control strategy are spliced ​​to construct value assessment data, and the value assessment data is sent to the value assessment network model for processing, and the strategy value data is output. The strategy gradient value of the strategy value data for the target headlight control strategy is calculated, and then the parameters of the headlight control strategy output model are adjusted using the gradient ascent method based on the strategy gradient value, so as to realize the real-time training of the headlight control strategy output model; at the same time, it also includes the real-time training of the value assessment network model; At the target network update time point, the headlight control strategy output model is directly used as the target headlight control strategy output model to replace the previous target headlight control strategy output model, and the time period between two adjacent target network update time points includes N monitoring time points, and the time period between two adjacent target network update time points is recorded as the analysis period; Real-time training of the value assessment network model includes the following steps: Obtain all value assessment data acquired in the current analysis cycle to form a local buffer value assessment database, and at the same time obtain the current geographic location of the vehicle, and match the cloud buffer value assessment database in the current analysis cycle from the cloud server based on the current geographic location of the vehicle; All the value assessment data in the local buffer value assessment database and the cloud buffer value assessment database within the current analysis cycle are sequentially input into the target value assessment network model for processing to obtain several target strategy value data, and the largest target strategy value data among all the target strategy value data is selected as the identification strategy value data. The loss value is calculated based on the identification strategy value data and the strategy value data, and the parameters of the value assessment network model are adjusted through the back propagation algorithm based on the loss value to realize real-time training of the value assessment network model.

2. The method for intelligently controlling LED vehicle lights based on environment perception according to claim 1, characterized in that: Based on the current geographic location of the vehicle, the cloud buffer value assessment database within the current analysis period is matched from the cloud server, specifically including the following steps: For any vehicle, all value assessment data acquired in the current analysis cycle are annotated by geographic location, cloud buffer value assessment data is constructed and uploaded to the cloud server; Calculate the distance between the vehicle's current geographic location and the geographic location corresponding to all cloud buffer value assessment data, record it as the regional distance, compare all regional distances with the preset distance, and form a cloud buffer value assessment database with the cloud buffer value assessment data corresponding to all regional distances less than the preset distance to complete the matching of the cloud buffer value assessment database.

3. The method for intelligently controlling LED vehicle lights based on environment perception according to claim 2, characterized in that: Training the output model of the vehicle light control strategy includes the following steps: A plurality of headlight control strategy output training samples are obtained, the headlight control strategy output training samples include environmental state data and its corresponding optimal headlight control strategy, all headlight control strategy output training samples are combined into a headlight control strategy output training set, and the headlight control strategy output training set is sent to a headlight control strategy output model with initialized parameters for training. During the training, the environmental state data is used as input, and the optimal headlight control strategy is used as a target. The first accuracy rate corresponding to the headlight control strategy output model is calculated, and it is determined whether the first accuracy rate corresponding to the headlight control strategy output model is higher than the accuracy threshold. If the first accuracy rate corresponding to the headlight control strategy output model is higher than the accuracy threshold, the trained headlight control strategy output model is output, and the initial headlight control strategy output model and the target headlight control strategy output model are both trained headlight control strategy output models; otherwise, the headlight control strategy output model is continuously trained through the headlight control strategy output training set.

4. The method for intelligently controlling LED vehicle lights based on environment perception according to claim 3 is characterized in that: The training of the value assessment network model includes the following steps: A number of positive value assessment training samples and negative value assessment training samples are obtained, wherein the positive value assessment training samples include environmental state data and its corresponding optimal headlight control strategy, and the negative value assessment training samples include environmental state data and its corresponding worst headlight control strategy. The positive value assessment training samples and the negative value assessment training samples are marked by the target value, and all the marked positive value assessment training samples and the negative value assessment training samples form a value assessment training set, and the value assessment training set is sent to the parameter-initialized value assessment network model for training. During the training, the positive value assessment training samples or the negative value assessment training samples are used as input, and the target value is used as the target. The second accuracy rate corresponding to the value assessment network model is calculated, and it is determined whether the second accuracy rate corresponding to the value assessment network model is higher than the accuracy threshold. If the second accuracy rate corresponding to the value assessment network model is higher than the accuracy threshold, the trained value assessment network model is output, and the initial value assessment network model and the target value assessment network model are both trained value assessment network models; otherwise, the value assessment network model is continuously trained by the value assessment training set; and the hyperparameters of the value assessment network model are adjusted by a genetic algorithm.

5. The method for intelligently controlling LED vehicle lights based on environment perception according to claim 4, characterized in that: The headlight control strategy output model and value assessment network model are both established based on the BP neural network model.

6. The method for intelligently controlling LED vehicle lights based on environment perception according to claim 5, characterized in that: The hyperparameters of the headlight control strategy output model and the hyperparameters of the value assessment network model are adjusted through genetic algorithms.

7. An intelligent control system for LED lights based on environmental perception, characterized in that: The system applies an LED vehicle light intelligent control method based on environment perception as described in any one of claims 1 to 6, including: An environmental status data acquisition module is used to acquire environmental status data during vehicle driving at a monitoring time point; The headlight control strategy output module is used to send the current environment status data to the headlight control strategy output model for processing and output the LED headlight control strategy; A headlight control module is used to intelligently control the LED headlights on the vehicle through an LED headlight control strategy; The real-time training module of the headlight control strategy output model is used to send the environmental state data to the target headlight control strategy output model for processing at the monitoring time point, output the target headlight control strategy, then splice the environmental state data with the target headlight control strategy, construct value assessment data, send the value assessment data to the value assessment network model for processing, output the strategy value data, calculate the strategy gradient value of the strategy value data for the target headlight control strategy, and then adjust the parameters of the headlight control strategy output model using the gradient ascent method based on the strategy gradient value, so as to realize the real-time training of the headlight control strategy output model; The value assessment network model real-time training module is used to perform real-time training on the value assessment network model.

Images