Vehicle lamp control system and method based on sparse gating hybrid expert model

Abstract

The invention provides a vehicle lamp control system and method based on a sparse gating hybrid expert model, and relates to the technical field of vehicle lamp adaptive control, and the method comprises the steps: obtaining ultrasonic, image and audio multi-mode signals, and carrying out the preprocessing; after the preprocessed multi-mode signals are input into the sparse gating hybrid expert model, ultrasonic, image and audio signals are respectively sent into an independent expert network, each expert network is responsible for analyzing input signals of a specific type, the ultrasonic expert network analyzes the ultrasonic signals and outputs the distance and the type of a front object, and the distance and the type of the front object are analyzed; the image expert network analyzes image features to obtain the distance and category of a front object and a preliminary vehicle lamp code, the audio expert network analyzes sound features to obtain the distance and category of the front object, and finally outputs of the two expert networks are selected through a routing network gating mechanism for weighted combination to generate a final vehicle lamp code. And obtaining a control signal of the vehicle lamp.

Description

Technical Field

[0001] The present disclosure relates to the technical field of vehicle light adaptive control, and in particular to a vehicle light control system and method based on a sparse gated hybrid expert model. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Although traditional car high beams can provide a longer field of vision when driving at night, they can also easily cause glare to oncoming vehicles, affecting the other driver's vision, thereby increasing the risk of traffic accidents. To solve this problem, the adaptive high beam system (ADB) came into being. The ADB system can dynamically adjust the illumination range of the high beam according to the road conditions ahead and the surrounding environment to avoid glare while providing sufficient lighting.

[0004] The existing ADB system mainly relies on a single camera signal to detect objects ahead and road conditions, but a single sensor signal often has limitations, such as low recognition accuracy under bad weather conditions, insufficient accuracy when detecting objects ahead, easy misjudgment or missed detection, and difficulty in fully and accurately identifying objects in various environments. In the existing multi-sensor fusion method, due to the large differences in data processing methods and algorithms of different sensors, the fusion is difficult. In addition, the existing ADB system has poor real-time performance, the existing system has a slow response speed, and cannot adjust the lights in time, affecting driving safety. The system also has a large amount of computational data processing, especially the complex image processing algorithm, which leads to high latency. The performance in complex environments (such as rainy and snowy weather, strong light interference, etc.) is unstable, prone to misjudgment, insufficient adaptability to environmental changes, and lacks effective robustness design. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a headlight control system and method based on a sparse gated hybrid expert model. Through the fusion of multi-sensor signals, based on a sparse gated hybrid expert model (SMoE), a router network is used to utilize linear layers and Top-K gating mechanisms to realize separate processing of multiple signals, and multiple final output results are combined into a final headlight code, thereby realizing high-precision perception of the front environment and intelligent control of the high beam.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions:

[0007] A vehicle light control method based on a sparse gated hybrid expert model includes:

[0008] Acquire ultrasonic, image and audio multimodal signals and pre-process them;

[0009] The preprocessed multimodal signal is input into the sparse gated hybrid expert model, and the headlight control signal is output to achieve adaptive adjustment of the headlight;

[0010] Among them, after the preprocessed multimodal signal is input into the sparse gated hybrid expert model, the ultrasonic, image and audio signals are sent to an independent expert network respectively. Each expert network is responsible for analyzing a specific type of input signal. The ultrasonic expert network analyzes the ultrasonic signal and outputs the distance and category of the object in front. The image expert network analyzes the image features to obtain the distance, category and preliminary headlight coding of the object in front. The audio expert network analyzes the sound features to obtain the distance and category of the object in front. Finally, the outputs of the two expert networks are selected through the routing network gating mechanism for weighted combination to generate the final headlight coding and obtain the headlight control signal.

[0011] According to some embodiments, the present disclosure adopts the following technical solutions:

[0012] The headlight control system based on sparse gated hybrid expert model includes:

[0013] Data acquisition module, used to acquire ultrasonic, image and audio multimodal signals and pre-process them;

[0014] A headlight control module is used to input the preprocessed multimodal signal into the sparse gated hybrid expert model, and output a headlight control signal to achieve adaptive adjustment of the headlight;

[0015] Among them, after the preprocessed multimodal signal is input into the sparse gated hybrid expert model, the ultrasonic, image and audio signals are sent to an independent expert network respectively. Each expert network is responsible for analyzing a specific type of input signal. The ultrasonic expert network analyzes the ultrasonic signal and outputs the distance and category of the object in front. The image expert network analyzes the image features to obtain the distance, category and preliminary headlight coding of the object in front. The audio expert network analyzes the sound features to obtain the distance and category of the object in front. Finally, the outputs of the two expert networks are selected through the routing network gating mechanism for weighted combination to generate the final headlight coding and obtain the headlight control signal.

[0016] According to some embodiments, the present disclosure adopts the following technical solutions:

[0017] A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the vehicle light control method based on a sparse gated hybrid expert model is implemented.

[0018] According to some embodiments, the present disclosure adopts the following technical solutions:

[0019] A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the vehicle light control method based on a sparse gated hybrid expert model is implemented.

[0020] According to some embodiments, the present disclosure adopts the following technical solutions:

[0021] An electronic device comprises: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes the vehicle light control method based on a sparse gated hybrid expert model.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The disclosed headlight control method based on sparse gated hybrid expert model and the disclosed adaptive high beam system can accurately detect the front object in various environments and intelligently adjust the illumination range and intensity of the headlights, significantly improving the safety and comfort of night driving. The disclosed method can improve safety by intelligently controlling the illumination range and intensity of the headlights, reducing interference to other road users and improving the safety of night driving; improve comfort by adjusting the headlights in real time according to changes in the front environment, improving the driver's field of vision and comfort.

[0024] The disclosed headlight control method based on sparse gated hybrid expert model improves the robustness and accuracy of the system by fusing signals from multiple sensors, adopts sparse hybrid expert model (SMoE) to analyze the world, increases the number of model parameters while maintaining low computational cost, and selects two expert networks to process each input signal, significantly improving the robustness and accuracy of the system while maintaining low computational cost. Experimental results show that the disclosed system can work effectively in various environments and significantly improves the safety and comfort of night driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0026] Figure 1 The figure is a schematic diagram of the method flow of an embodiment of the present disclosure. DETAILED DESCRIPTION

[0027] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0029] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0030] Example 1

[0031] In one embodiment of the present disclosure, a vehicle light control method based on a sparse gated hybrid expert model is provided, comprising the following steps:

[0032] Step 1: Acquire ultrasonic, image and audio multimodal signals and pre-process them;

[0033] Step 2: Input the preprocessed multimodal signal into the sparse gated hybrid expert model, and output the headlight control signal to achieve adaptive adjustment of the headlight;

[0034] Among them, after the preprocessed multimodal signal is input into the sparse gated hybrid expert model, the ultrasonic, image and audio signals are sent to an independent expert network respectively. Each expert network is responsible for analyzing a specific type of input signal. The ultrasonic expert network analyzes the ultrasonic signal and outputs the distance and category of the object in front. The image expert network analyzes the image features to obtain the distance, category and preliminary headlight coding of the object in front. The audio expert network analyzes the sound features to obtain the distance and category of the object in front. Finally, the outputs of the two expert networks are selected through the routing network gating mechanism for weighted combination to generate the final headlight coding and obtain the headlight control signal.

[0035] As an embodiment, the specific implementation process of a vehicle light control method based on a sparse gated hybrid expert model disclosed in the present invention is as follows:

[0036] Step 1: Acquire ultrasonic, image and audio multimodal signals and preprocess them;

[0037] Specifically, the ultrasonic signal is collected by an ultrasonic sensor, and the ultrasonic signal reflected by the front object is obtained by emitting the ultrasonic signal; the image signal is collected by a camera, and the audio signal is collected by a microphone;

[0038] The preprocessing method is signal normalization and denoising operation.

[0039] Step 2: Input the preprocessed multimodal signal into the sparse gated hybrid expert model, and output the headlight control signal to achieve adaptive adjustment of the headlight;

[0040] Specifically, the sparse gated hybrid expert model is based on the transformer architecture and uses a sparse hybrid expert layer, which is composed of a combination of multiple expert network sub-models. Each expert network processes a specific type of input signal. The sparse hybrid expert layer is followed by a routing network, which is a trainable gated network used to select two expert networks to weightedly process the current input signal and output a sparse n-dimensional vector. It includes:

[0041] The normalized ultrasonic echo features, image feature matrix, audio, lidar signal features, infrared signals and other features are cross-modally embedded and aligned, and the temporal correlation is established through the position encoding layer. The multi-head self-attention mechanism is used to capture the dynamic interaction between cross-modal features, and the pre-processed ultrasonic, image and audio signals are input into the expert network respectively. The expert specializes in processing specific signals. The expert contains 3 layers of gated recurrent units (trainable gated networks designed based on transformer architecture, and the input is multi-modal feature concatenation vectors), equipped with independent feature extraction branches to calculate the cosine similarity between the input signal and each expert network, and retain the corresponding Top experts.

[0042] Furthermore, the results of the outputs processed by each expert network are input into the routing network. The routing network selects two expert networks to process the current input signals and weightedly outputs the final headlight control signal. The generated headlight control signal is sent to the headlight control system. The control system adjusts the illumination range and intensity of the headlight according to the headlight control signal to realize the function of adaptively adjusting the high beam.

[0043] Specifically, a three-layer gated recurrent unit (a trainable gated network designed based on the transformer architecture, with the input being a multimodal feature concatenation vector) is equipped with an independent feature extraction branch to calculate the cosine similarity between the input signal and each expert network, and the corresponding Top experts are retained.

[0044] The KL divergence evaluation is performed on the top experts, and finally the two best experts are selected to output weight vectors for Softmax normalization. After the selected two experts process the input features respectively, the features are spliced ​​according to the weights:

[0045] F_fusion=w_1·Expert_1(X)⊕w_2·Expert_2(X)

[0046] Among them, ⊕ represents the channel dimension splicing operation, which retains the independence of the output features of each expert (using weighted summation or attention mechanism to fuse the output features of two experts).

[0047] As an embodiment, in the headlight control method based on the sparse gated hybrid expert model disclosed in the present invention, the Moe system can process signals from different sensors, including ultrasonic signals, camera signals and audio signals. The ultrasonic sensor provides distance and category information, the camera provides image information and generates preliminary headlight coding through the Transformer layer, and the audio signal provides environmental sound information to assist the Moe system to better output the final headlight coding. The Moe system includes multiple expert networks, each of which is responsible for processing a specific type of input signal. The appropriate expert network is selected through the router network to process the current input signal, ensuring that each input signal can be processed by the most appropriate expert network, thereby improving the accuracy and real-time performance of the system. The Moe system adopts a sparse gating mechanism, and each input signal only activates a few expert networks, not all. The Moe system ensures load balancing of each expert network through a dynamic load balancing mechanism. By introducing a load balancing loss function, the Moe system can effectively allocate computing resources, avoid overloading of certain expert networks, and improve the overall performance of the system. By training the Moe system, it can automatically select the most suitable expert network according to the characteristics of the input signal, thereby improving the robustness and adaptability of the system.

[0048] Among them, the ultrasonic sensor is a non-contact measurement tool that measures the distance of the object in front by emitting and receiving ultrasonic signals. The ultrasonic sensor used in the present disclosure has been trained to output the distance to the object and the category of the object like the brain of a bat. Specifically, the input of the ultrasonic model is the ultrasonic signal, and the output is the distance to the object and the category of the object. The camera is used to capture the image of the environment in front and output the distance to the object and the category of the object through the trained neural network model. In addition, after the camera signal is processed by the Transformer layer, it can output the headlight code.

[0049] The combination of a mixed expert model (Mixture of Experts, MoE) is a neural network architecture that processes input signals through a combination of multiple expert networks (i.e., sub-models). Each expert network is responsible for processing a specific type of input signal, and selects the outputs of two expert networks through a routing network for weighted combination to finally generate a headlight control signal. This architecture can significantly increase the number of parameters of the model while maintaining a low computational cost. The mixed expert model disclosed in the present invention is based on a transformer architecture (Transformer architecture) and uses a sparse mixed expert layer (Sparse Mixture of Experts, SMoE). Routing network The routing network is a trainable gating network that selects two expert networks to process the current input signal. The output of the routing network is a sparse n-dimensional vector, where n is the number of expert networks. By using a sparse mixed expert model, the present disclosure can significantly increase the number of parameters of the model while maintaining a low computational cost. Specifically, each input signal only needs to process two expert networks, so the number of parameters actually used for each input signal is about 1.3 billion, while the total number of parameters is 4.7 billion. This architecture enables the model to perform well on large-scale data sets while maintaining high computational efficiency. The collected data is first preprocessed, including normalization, denoising and other operations to ensure the quality of the input signal. The preprocessed signal is sent to the corresponding neural network model for processing: the ultrasonic signal is processed by the ultrasonic model and outputs the distance to the object and the category of the object. The camera signal is processed by the camera model and outputs the distance to the object and the category of the object, as well as the light code.

[0050] As an embodiment, laser radar signals and infrared signals can also be obtained. The laser radar signal is processed by the laser radar model to output the distance to the object and the category of the object, which is used to model the three-dimensional space; the infrared signal is processed by the infrared model to determine whether there are pedestrians; the processed signal is sent to the routing network, and the routing network selects two expert networks to process the current input signal. The generated headlight control signal is sent to the headlight control system, and the control system adjusts the illumination range and intensity of the headlight according to the signal, thereby realizing the function of adaptive high beam.

[0051] The ultrasonic distance measurement disclosed in the present invention is uploaded to the server after the measurement is completed in the vehicle body domain control. The measurement formula is as follows:

[0052]

[0053] Where d is the distance, c is the speed of sound (343m / s), and t is the round trip time of the ultrasonic signal.

[0054] The Adaptive Driving Beam (ADB) system is designed to improve the safety and comfort of nighttime driving by intelligently sensing the surrounding environment and automatically adjusting the vehicle's high beam. The purpose of the Adaptive Driving Beam (ADB) system is to prevent dazzling other drivers. The Adaptive Driving Beam system automatically detects obstacles and vehicles ahead and adjusts the illumination range of the lights to ensure the best lighting effect without affecting other drivers.

[0055] The present disclosure provides an adaptive high beam system, which specifically includes the following components: an ultrasonic sensor for detecting the distance and category of a front object. The working principle of the ultrasonic sensor is similar to the echolocation of a bat, which calculates the distance by emitting ultrasonic waves and receiving reflected waves. The ultrasonic model is trained to output the distance to the object and the category of the object. A camera is used to capture image information in front. The camera model is trained to output the distance to the object and the category of the object, and generates the headlight code through the Transformer layer.

[0056] The ultrasonic sensors are installed on the left and right sides of the front of the vehicle and connected to the main control unit through the CAN bus. The camera is installed on the front of the vehicle and connected to the main control unit through the high-speed data interface Ethernet protocol. The headlight controller receives the control signal of the main control unit and adjusts the illumination range and intensity of the high beam. The ultrasonic sensor transmits ultrasonic signals, receives reflected waves, and calculates the distance. At the same time, the reflected ultrasonic signals are transmitted to the cloud, and the ultrasonic model in the cloud processes the signals and outputs the distance to the object and the category of the object. The camera captures the front image, and the camera model outputs the distance to the object and the category of the object based on the image content. Through the Transformer layer, this information is converted into the headlight code. The headlight controller adjusts the illumination range and intensity of the high beam according to the headlight code generated by the main control unit. The training process of the ultrasonic model first requires data collection to collect ultrasonic signals of the front environment through ultrasonic sensors. Data preprocessing normalizes and denoises the collected signals. Model training uses the training data set to train the ultrasonic model. The Adam optimizer is used during the training process, and the initial learning rate increases linearly and then gradually decreases. Model evaluation evaluates the performance of the model on the test data set to ensure that the model can accurately output the distance to the object and the category of the object. The experimental results show that the model has learned that different objects have different reflections, and uses their reflection, Doppler effect, transmission, etc. to obtain information measurement objects and processing. The training process of the camera model also starts with data collection, which collects images of the environment in front through the camera.

[0057] Data preprocessing normalizes and denoises the collected images, and passes the signals to the Moe model for training. Model evaluation evaluates the performance of the model on the test dataset to ensure that the model can accurately output the distance to the object, the category of the object, and the headlight coding.

[0058] As an embodiment, the mapping strategy of the headlight position encoding can generally be adjusted according to the visual information features and the position information of the target detection to establish a mapping layer from the image coordinate system to the headlight control signal. A binary search algorithm is used to quickly locate the target coordinates in space. There are two formats for the output of the target detection algorithm, the coco format and the voc format, in which the voc format can directly use the binary search method to directly locate the center coordinate information of the target after restoring the coordinate information. The four values ​​(x1, x2, y1, y2) output in the voc format are directly used ((x1+x2) / 2, (y1+y2) / 2). Because the voc format outputs the normalized values, to locate the center point of the target detection, it is necessary to change it to the original coordinate position, determine the center position of the target detection, and determine whether it is located on the left or right side of the car according to the position. Because the high beam LED generally uses 8 LEDs, the interval can be determined according to the number of LEDs used in different models. Knowing the center position of the target detection, the binary search algorithm can be used to determine the interval where the center position of the target is located. The adaptive high beam ADB determines the LED that needs to be extinguished according to the interval. The output result is in coco format, then y2=y+h, x2=x+w, determine the target center position according to the above scheme, and repeat the above algorithm. Of course, the headlight coding layer can also be specified by reinforcement learning, but the problem is the accuracy of the headlight coding layer output by reinforcement learning. The safety of deep learning and reinforcement learning on the car computer is questionable, so this patent still uses the traditional manual program to set the headlight coding layer when designing the headlight coding layer, including the control of the motor, and the output results of the neural network are output after being judged by the program.

[0059] Experiments were conducted in a variety of scenarios, including urban roads, highways, and rural roads. The experimental results show that the adaptive high beam system of the present invention can accurately detect objects ahead in various environments, and intelligently adjust the illumination range and intensity of the headlights, significantly improving the safety and comfort of night driving. This patent can improve safety, reduce interference to other road users, and improve the safety of night driving by intelligently controlling the illumination range and intensity of the headlights. Improve comfort Adjust the headlights in real time according to changes in the environment ahead to improve the driver's field of vision and comfort. Multimodal fusion improves the robustness and accuracy of the system by fusing signals from multiple sensors. Computational efficiency uses a sparse mixture of experts model (SMoE) to maintain a low computational cost while increasing the number of model parameters.

[0060] Example 2

[0061] In one embodiment of the present disclosure, a vehicle light control system based on a sparse gated hybrid expert model is provided, comprising:

[0062] Data acquisition module, used to acquire ultrasonic, image and audio multimodal signals and pre-process them;

[0063] A headlight control module is used to input the preprocessed multimodal signal into the sparse gated hybrid expert model, and output a headlight control signal to achieve adaptive adjustment of the headlight;

[0064] Among them, after the preprocessed multimodal signal is input into the sparse gated hybrid expert model, the ultrasonic, image and audio signals are sent to an independent expert network respectively. Each expert network is responsible for analyzing a specific type of input signal. The ultrasonic expert network analyzes the ultrasonic signal and outputs the distance and category of the object in front. The image expert network analyzes the image features to obtain the distance, category and preliminary headlight coding of the object in front. The audio expert network analyzes the sound features to obtain the distance and category of the object in front. Finally, the outputs of the two expert networks are selected through the routing network gating mechanism for weighted combination to generate the final headlight coding and obtain the headlight control signal.

[0065] Furthermore, the sparse gated hybrid expert model is based on a transformer architecture and uses a sparse hybrid expert layer, which is composed of a combination of multiple expert network sub-models. Each expert network processes a specific type of input signal. The sparse hybrid expert layer is followed by a routing network, which is a trainable gated network for selecting two expert networks to weightedly process the current input signal and output a sparse n-dimensional vector. The results of the processing outputs of each expert network are input to the routing network, which selects two expert networks to process the current input signal and weightedly outputs the final headlight control signal. The generated headlight control signal is sent to the headlight control system, and the control system adjusts the illumination range and intensity of the headlight according to the headlight control signal to achieve the function of adaptively adjusting the high beam.

[0066] Example 3

[0067] In one embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the vehicle light control method based on the sparse gated hybrid expert model is implemented.

[0068] Example 4

[0069] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the vehicle light control method based on the sparse gated hybrid expert model is implemented.

[0070] Example 5

[0071] In one embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the vehicle light control method based on the sparse gated hybrid expert model.

[0072] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0074] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A vehicle light control method based on a sparse gated hybrid expert model, characterized in that: include: Acquire ultrasonic, image and audio multimodal signals and pre-process them; The preprocessed multimodal signal is input into the sparse gated hybrid expert model, and the headlight control signal is output to achieve adaptive adjustment of the headlight; Among them, after the preprocessed multimodal signal is input into the sparse gated hybrid expert model, the ultrasonic, image and audio signals are sent to an independent expert network respectively. Each expert network is responsible for analyzing a specific type of input signal. The ultrasonic expert network analyzes the ultrasonic signal and outputs the distance and category of the object in front. The image expert network analyzes the image features to obtain the distance, category and preliminary headlight coding of the object in front. The audio expert network analyzes the sound features to obtain the distance and category of the object in front. Finally, the outputs of the two expert networks are selected through the routing network gating mechanism for weighted combination to generate the final headlight coding and obtain the headlight control signal.

2. The vehicle light control method based on sparse gated hybrid expert model according to claim 1, characterized in that: The ultrasonic signal is collected by an ultrasonic sensor, and the ultrasonic signal reflected by the front object is obtained by emitting the ultrasonic signal; the image signal is collected by a camera, and the audio signal is collected by a microphone. The preprocessing method is signal normalization and denoising operation.

3. The vehicle light control method based on sparse gated hybrid expert model according to claim 1, characterized in that: The sparse gated hybrid expert model is based on a transformer architecture and uses a sparse hybrid expert layer. The sparse hybrid expert layer is composed of a combination of multiple expert network sub-models. Each expert network processes a specific type of input signal. The sparse hybrid expert layer is followed by a routing network. The routing network is a trainable gating network that is used to select two expert networks to weightedly process the current input signal and output a sparse n-dimensional vector.

4. The vehicle light control method based on sparse gated hybrid expert model according to claim 1, characterized in that: The results of the outputs processed by each expert network are input into the routing network. The routing network selects two expert networks to process the current input signals and weightedly outputs the final headlight control signal. The generated headlight control signal is sent to the headlight control system. The control system adjusts the illumination range and intensity of the headlight according to the headlight control signal to realize the function of adaptively adjusting the high beam.

5. A headlight control system based on a sparse gated hybrid expert model, characterized in that: include: Data acquisition module, used to acquire ultrasonic, image and audio multimodal signals and pre-process them; A headlight control module is used to input the preprocessed multimodal signal into the sparse gated hybrid expert model, and output a headlight control signal to achieve adaptive adjustment of the headlight; Among them, after the preprocessed multimodal signal is input into the sparse gated hybrid expert model, the ultrasonic, image and audio signals are sent to an independent expert network respectively. Each expert network is responsible for analyzing a specific type of input signal. The ultrasonic expert network analyzes the ultrasonic signal and outputs the distance and category of the object in front. The image expert network analyzes the image features to obtain the distance, category and preliminary headlight coding of the object in front. The audio expert network analyzes the sound features to obtain the distance and category of the object in front. Finally, the outputs of the two expert networks are selected through the routing network gating mechanism for weighted combination to generate the final headlight coding and obtain the headlight control signal.

6. The vehicle light control system based on sparse gated hybrid expert model as claimed in claim 5, characterized in that: The sparse gated hybrid expert model is based on a transformer architecture and uses a sparse hybrid expert layer. The sparse hybrid expert layer is composed of a combination of multiple expert network sub-models. Each expert network processes a specific type of input signal. The sparse hybrid expert layer is followed by a routing network. The routing network is a trainable gating network that is used to select two expert networks to weightedly process the current input signal and output a sparse n-dimensional vector.

7. The vehicle light control system based on sparse gated hybrid expert model as claimed in claim 5, characterized in that: The results of the outputs processed by each expert network are input into the routing network. The routing network selects two expert networks to process the current input signals and weightedly outputs the final headlight control signal. The generated headlight control signal is sent to the headlight control system. The control system adjusts the illumination range and intensity of the headlight according to the headlight control signal to realize the function of adaptively adjusting the high beam.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the vehicle light control method based on a sparse gated hybrid expert model according to any one of claims 1 to 4 is implemented.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the vehicle light control method based on the sparse gated hybrid expert model as described in any one of claims 1 to 4 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes the vehicle light control method based on the sparse gated hybrid expert model as described in any one of claims 1 to 4.

Images