Dynamic glare suppression method and system based on multi-vehicle cooperative perception, vehicle and computer readable storage medium

Abstract

The invention belongs to the technical field of intelligent automobile lighting, and particularly relates to a dynamic glare suppression method and system based on multi-automobile cooperative perception, a vehicle and a computer readable storage medium. According to the method, a glare sensitivity distribution diagram which is sent by an opposite vehicle and based on the visual angle of the vehicle is received through V2X vehicle wireless communication, vehicle-vehicle information sharing is achieved, and the sensing limitation of a single vehicle is broken through; constructing a human eye glare sensitivity model based on the CIE standard, converting the physical brightness into subjective glare perception to generate a smooth weight map, and performing curve and ramp dynamic geometric compensation on the map to eliminate perspective distortion; and finally, generating a pixel-level dimming instruction by using the multi-agent depth deterministic strategy gradient network in combination with the lighting requirement of the vehicle, and driving the pixelated headlight to execute accurate brightness adjustment. According to the method, the problems of multi-vehicle meeting game conflicts, rough control strategies and failure of complex road conditions are effectively solved, and collaborative and accurate glare suppression and remarkable improvement of subjective comfort are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent car lights, and in particular to a dynamic glare suppression method and system based on multi-vehicle cooperative perception, a vehicle and a computer readable storage medium. BACKGROUND

[0002] With the continuous development of automotive lighting technology, adaptive driving beam (ADB) has become an important part of intelligent vehicles because it can effectively improve the night driving field of view while avoiding glare to other road users. Currently, ADB technology mainly relies on mechanical baffles, digital light processing technology or emerging microLED pixel arrays to achieve regional shielding or adjustment of high beam light. However, the core control of existing ADB systems still stays at the level of single-vehicle intelligence, i.e., relying solely on the vehicle's sensors (such as cameras, millimeter wave radars) to perceive the environment and make independent decisions. This inherent technical architecture leads to a series of technical defects in complex real-world scenarios: At the perception level, single-vehicle perception has limitations and high false detection rates. Due to the reliance on a single perspective of the vehicle's sensors, the system has difficulty accurately identifying and locating oncoming vehicles in unstructured road scenarios such as curves and slopes with perspective distortion, resulting in the generated dark areas being offset or missing shielding, which causes the glare suppression function to fail; At the control strategy level, existing systems lack multi-vehicle coordination and fine-tuning capabilities. In multi-vehicle meeting scenarios, each vehicle makes independent light adjustment decisions, lacks information exchange and collaborative optimization mechanisms between vehicles, and is prone to glare tug-of-war phenomena, making it impossible to achieve optimal glare suppression for the entire system. Even with microLED headlights that can control at the pixel level, existing control strategies are limited to simple binary processing of "bright" and "dark", and fail to integrate the visual sensitivity characteristics of the human eye for dynamic and smooth weight adjustment at the software level, resulting in poor subjective visual comfort; In addition, at the system robustness level, existing technologies lack consideration of forward-looking information and degradation strategies. Due to the lack of support for V2X vehicle wireless communication technology, the system cannot obtain accurate trajectory and state information of oncoming vehicles in advance, limiting its predictive ability. At the same time, once the main sensor is disturbed or fails, the entire glare suppression function may completely degrade, lacking effective fallback strategies.

[0003] Therefore, the existing adaptive driving beam system has obvious deficiencies in perception accuracy, multi-vehicle coordination, control refinement, and system robustness. Therefore, there is an urgent need in the art for a new glare suppression solution that can overcome the limitations of single-vehicle perception, achieve vehicle-to-vehicle coordination, and integrate human factors engineering for pixel-level fine-tuning. SUMMARY

[0004] The technical problems solved by the present application are: in order to solve the problems of the prior art in the background art, such as the adaptive high beam system in the prior art, the sensing accuracy, multi-vehicle cooperation, control refinement and system robustness are all obviously insufficient, a dynamic glare suppression method, system, vehicle and computer readable storage medium based on multi-vehicle cooperative sensing are provided.

[0005] The technical solution adopted by the present application to solve its technical problems is: a dynamic glare suppression method based on multi-vehicle cooperative sensing, comprising the following steps: S1, acquiring multi-source sensing information, at least including receiving sensing information sent by at least one oncoming vehicle through vehicle wireless communication V2X; the sensing information sent by the oncoming vehicle includes a glare sensitivity distribution map based on the object vehicle perspective; S2, constructing a human eye glare sensitivity model, processing the glare sensitivity distribution map, and generating a smooth weight map for dimming control; S3, performing geometric compensation on the glare sensitivity distribution map; S4, based on the smooth weight map after S2 compensation and the self-vehicle road lighting demand, generating dimming instructions for controlling the brightness of each pixel of the pixelated vehicle lamp through an optimization algorithm; S5, executing the dimming instructions in S4 to suppress the glare of the oncoming vehicle driver.

[0006] By introducing V2X communication to obtain the sensing information of the oncoming vehicle in S1, the limitation of single-vehicle sensing is broken through, the large recognition error and dark area offset caused by perspective distortion are solved from the source, and the foundation for multi-vehicle cooperation is laid; S2 constructs a human eye glare model, which improves the control target from simple physical light shielding to intelligent dimming based on human eye comfort, and solves the problem of rough control strategy; The geometric compensation of S3 is specially optimized for curved road / slope scene, which further improves the suppression accuracy under complex road conditions; The optimization algorithm of S4 generates the optimal instruction by integrating the cooperative information and the visual model.

[0007] Further, the step S1 is specifically: S11, real-time pulling the longitudinal speed, yaw angular speed, steering wheel angle and steering light state of the vehicle through the CAN bus, and constructing a vehicle motion trajectory prediction model; S12, receiving the CAM message sent by the oncoming vehicle through the C-V2X vehicle-mounted unit, analyzing the 256*64 grayscale map carried therein, and obtaining the glare sensitivity distribution map; When there are multiple object vehicles, different fusion weights are assigned to the glare sensitivity distribution maps sent by each oncoming vehicle according to the distance between the ego vehicle and each oncoming vehicle , and the fusion weights are based on Linearly superimpose multiple glare sensitivity distribution maps to generate a unified total sensitivity distribution map : , wherein, , is the longitudinal distance between the vehicle and the kth oncoming vehicle; is the most appropriate pixel grayscale map given for each oncoming vehicle; S13, the front camera of the vehicle continuously outputs the oncoming vehicle contour, lane line and road slope angle, and constructs a three-dimensional road coordinate system according to the lane line cubic polynomial and slope; S14, the millimeter wave radar synchronously gives the distance and speed redundancy, and uses the cascade matching and dynamic weight fusion algorithm to do cross verification with the radar point cloud, to ensure that the sensitive grayscale map corresponds to the real oncoming vehicle.

[0008] Through the distance weight fusion algorithm in S12, when there are multiple oncoming vehicles, the system can intelligently integrate the glare information of all vehicles to generate a more accurate and reliable total sensitivity map, thereby effectively eliminating the multi-car game conflict and avoiding the glare tug-of-war phenomenon; at the same time, through the radar and visual cross verification in S14, the false detection rate of oncoming vehicles is greatly reduced, ensuring that the cooperative information is correctly applied to the real vehicle, avoiding false occlusion, and realizing effective fusion and credibility enhancement of multi-car information.

[0009] Further, the S13 specifically comprises: S131, convert the image coordinates to the road local coordinates, and the calculation formula is: , wherein (u, v) is the pixel coordinates of the center of the oncoming vehicle bounding box in the image, K is the camera intrinsic matrix, and d is the actual distance corresponding to the pixel; , ) is the converted road local coordinates; S132, compensate for the slope and curvature, and the calculation formula is: , wherein, is the camera installation height, and a is the road slope angle; S133, map the two-dimensional sensitivity map to the three-dimensional road, and the calculation formula is: .

[0010] By converting two-dimensional image coordinates to three-dimensional road surface coordinates that better reflect the geometric relationship of the real world, it is prepared to correct the perception error caused by changes in vehicle posture and road shape, and to solve the problem of failure of glare suppression on curved roads / slopes.

[0011] Further, the S2 is based on a human eye visual sensitivity model, specifically: S21, based on the glare evaluation standard of the International Commission on Illumination CIE, the subjective glare value TI caused by a single pixel in the pixelized car light to the human eye is calculated: , Wherein, is the equivalent screen brightness, is the background adaptive brightness, and θ is the angle between the glare source and the line of sight; 65 is the scale constant, and 0.25 is the equivalent screen brightness contribution weight of effective adaptive brightness, 0.8 is the nonlinear perception correction index; S22, the subjective glare value distribution map obtained by calculation is smoothed using a Gaussian kernel to generate a smoothing weight map.

[0012] By introducing the internationally authoritative CIE glare standard formula, the adjustment of pixel-level brightness is no longer arbitrary, but is based on the human eye visual sensitivity, thereby realizing dynamic and fine adjustment at the software level, and significantly improving the subjective comfort of glare prevention.

[0013] Further, the compensation processing of S3 is specifically: S31, the lane line is fitted by a cubic polynomial, and the lane line curvature and slope angle are calculated in real time: S311, the lane line is fitted by a cubic polynomial: , x is the longitudinal distance in the vehicle forward direction, and y is the transverse distance in the vehicle forward direction; S312, the curvature calculation formula is obtained based on the standard definition formula for calculating the curvature of a two-dimensional plane curve in differential geometry: , Wherein, is the first derivative, which represents the slope of the lane line, is the second inverse, which represents the bending degree of the lane line; S313, the slope angle calculation formula is obtained based on the geometric slope definition and the numerical difference method: , Wherein, is a small forward distance interval; S32, a 3x3 perspective transformation matrix is generated according to the curvature and slope; S33, using bilinear interpolation for pixel-level resampling, specifically: S331, positioning four domain points:

[0014] S332, twice linear difference in X direction: At , according to the pixel values of and , the value of point is calculated by interpolation: , At , according to the pixel values of and , the value of point is calculated by interpolation: ; S333, linear interpolation in Y direction: Using the values of R1 and R2, interpolation is performed in the y direction to obtain the pixel value of the final target point : .

[0015] Through the calculation of S31 to S33, the dynamic perspective transformation matrix is generated and the image is resampled, which can make the system prediction dark area follow the eye position of the oncoming driver, completely solve the dark area deviation and light leakage problem in the scene of curve and slope.

[0016] Further, the light adjustment instruction generated by the optimization algorithm in S4 is specifically: The smooth weight map and the self-vehicle road lighting demand map are constructed as a joint state, and the calculation formula is: ; Wherein, is the glare sensitivity of the oncoming vehicle, is the self-vehicle road lighting demand, R is the risk weighting coefficient, and 256x64x2 is the joint state constructed according to the pixel resolution of 256x64 and the channel number of 2; The joint state is input into an offline trained multi-agent deep deterministic policy gradient MADDPG network, and the pixel-level pulse width modulation PWM duty cycle instruction output by the network is used as the light adjustment instruction.

[0017] Through the MADDPG reinforcement learning algorithm, the system can learn the optimal cooperative strategy in a complex multi-vehicle scene, and the output result is no longer a simple binary bright and dark, but a smooth pixel-level brightness instruction considering the needs of multiple parties; At the same time, the problem of rough multi-vehicle cooperative optimization and control strategy is solved, and the optimal overall performance of the system is realized.

[0018] Further, it further comprises a degradation step S6, specifically: Real-time monitoring of V2X communication link state; When monitoring the V2X communication interruption, enabling a visual perception model based on a lightweight neural network to estimate an alternative glare sensitivity distribution map according to the images collected by the camera of the ego vehicle; Based on the alternative glare sensitivity distribution map, steps S2 to S5 are performed.

[0019] By enabling a lightweight neural network for visual backup, the system can still maintain basic glare suppression performance after the main cooperative channel is interrupted, rather than completely degenerating, solving the problem of functional vulnerability caused by the lack of degradation strategy, and ensuring the availability and robustness when V2X communication fails.

[0020] Also provided is a suppression system of the dynamic glare suppression method based on multi-vehicle cooperative perception according to the above-mentioned scheme, comprising: A V2X communication module for receiving perception information containing a glare sensitivity distribution map sent to a vehicle in front; A processing unit for: a. Running a human eye visual sensitivity model to process the glare sensitivity distribution map to generate a smooth weight map; b. Running an optimization algorithm to generate a pixel-level dimming instruction based on the smooth weight map and the ego vehicle's road lighting needs; c. Based on the lane lines and slope information obtained by the ego vehicle sensors, performing a dynamic perspective transformation to correct the glare sensitivity distribution map; d. When the V2X communication module is interrupted, call the preset lightweight neural network to generate an alternative glare sensitivity distribution map; A pixelated car light driving module for receiving the dimming instruction and driving the pixelated car light to perform corresponding brightness changes.

[0021] Through the system architecture composed of the V2X communication module, the processing unit and the driving module, it is ensured that all method steps can be effectively executed, thereby providing the vehicle with a dynamic glare suppression capability with cooperative perception, humanized dimming and full-condition robustness as a whole.

[0022] Also provided is a vehicle equipped with a dynamic glare suppression system based on multi-vehicle cooperative perception as described in the above-mentioned scheme.

[0023] Also provided is a computer readable storage medium having stored thereon a computer program, the program being executed by a processor to implement the steps of the multi-vehicle cooperative perception based dynamic glare suppression method according to the above-mentioned solution.

[0024] Advantages of the present application: (1) The present application directly obtains a high-pixel resolution glare sensitivity distribution map from an oncoming vehicle through V2X communication, fundamentally breaking through the visual field limitation of single-vehicle perception. Through multi-source information fusion based on distance weight, the system can accurately grasp the glare sensitive area of all related road users, thereby realizing cooperative decision-making and overall optimization among multiple vehicles. The game conflict of "you dark and I bright" and the "glare tug-of-war" phenomenon caused by independent dimming of each vehicle in traditional systems are effectively eliminated, and the overall road glare intensity is minimized from a system level; (2) The international standard is introduced to construct a single-pixel glare evaluation model. Through the model, the physical brightness is converted into the subjective glare perception of the human eye, and a smooth weight map is generated, which drives the MADDPG reinforcement learning network to generate sub-centimeter precision dimming instructions. The light control strategy is upgraded from rough binary processing to dynamic and smooth adjustment considering the human eye visual sensitivity, greatly improving the subjective visual comfort during anti-glare process; (3) A dynamic three-dimensional road surface coordinate system is constructed by fusing lane line cubic polynomial fitting and real-time slope compensation. The dynamic perspective transformation matrix generated thereby is used to perform back projection and resampling on the sensitivity distribution map, which can accurately compensate for the perspective distortion caused by changes in road geometry. This significantly reduces the horizontal positioning error in dark areas, ensuring that the anti-glare area can accurately cover the oncoming driver on curves and slopes, effectively avoiding light leakage or false shielding; (4) A degradation mode is set. Once communication failure is monitored, the system immediately switches seamlessly to a visual perception backup scheme based on MobileNetV3-Small lightweight neural network, which estimates the sensitivity distribution by analyzing the vehicle camera image, thereby maintaining the core glare suppression performance in single-vehicle mode. At the same time, the system will trigger a fault code warning simultaneously, reminding the driver of the system state change, to ensure continuity and reliability under different working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0025] The present application will be further described below in conjunction with the drawings and examples.

[0026] Figure 1 is a flowchart of the multi-vehicle cooperative perception based dynamic glare suppression method of the present application.

[0027] Figure 2 is a flowchart of the multi-vehicle cooperative perception based dynamic glare suppression method of the present application using bilinear interpolation for pixel-level resampling.

[0028] Figure 3 is the control block diagram of the dynamic glare suppression system based on multi-vehicle cooperative perception of the present application. DETAILED DESCRIPTION

[0029] The present application will now be further described in greater detail in connection with the accompanying drawings. These drawings are not to scale and, with the exception of the results figures, are purely schematic and show only those parts of the application that are necessary in order to elucidate the application. The drawings in which:

[0030] As shown in the figure, a dynamic glare suppression method based on multi-vehicle cooperative perception is executed in 50ms fixed beat cycle, and the 256x64 pixel level PWM instruction of the lock is written into the microLED drive register in real time. The specific steps are as follows: Figure 1 Step one, get multi-source perception information: Pull the longitudinal speed, yaw rate, steering wheel angle and turn signal state of the vehicle in real time through CAN bus, and construct the vehicle motion trajectory prediction model; the purpose is to accurately predict the lane where the vehicle will appear next and how the vehicle head changes in the subsequent steps, so as to lock the possible visual field area of the opposite driver in advance. Receive the CAM message sent by the opposite vehicle through the C-V2X vehicle-mounted unit, analyze the 256x64 gray scale diagram carried in it, and get the glare sensitivity distribution diagram; after getting this diagram, the system knows the specific pixel position that the opposite driver is most afraid of being illuminated, which provides direct coordinate basis for subsequent accurate light-out; when there are multiple object vehicles, according to the distance between the vehicle and each opposite vehicle, different fusion weights are assigned to the glare sensitivity distribution diagrams sent by each opposite vehicle

[0031] , and based on the fusion weights , the multiple glare sensitivity distribution diagrams are linearly superimposed to generate a unified total sensitivity distribution diagram : , Among them, , is the longitudinal distance between the vehicle and the kth opposite vehicle; is the most appropriate pixel gray scale diagram for each opposite vehicle.

[0032] The front camera of the vehicle continuously outputs the outline of the opposite vehicle, the lane line and the road slope angle, and constructs a three-dimensional road coordinate system according to the lane line cubic polynomial and the slope, which is: Convert the image coordinates to local road coordinates, and the calculation formula is: , ​Where (u, v) is the pixel coordinate of the center of the bounding box of the oncoming vehicle in the image, K is the camera intrinsic matrix, and d is the actual distance corresponding to the pixel. , is the transformed local coordinate of the road surface; Then, the slope and curvature are compensated, and the calculation formula is: , Where, is the installation height of the camera, and a is the road slope angle. Next, the two-dimensional sensitivity map is mapped to the three-dimensional road surface, and the calculation formula is: .

[0033] The millimeter wave radar synchronously provides distance and speed redundancy, and the radar point cloud uses cascade matching and dynamic weight fusion algorithm for cross verification to ensure that the sensitive grayscale map corresponds to the real oncoming vehicle.

[0034] Step two, build a human eye glare sensitivity model: The single-pixel glare formula in the CIE 150-2017 standard is moved into the ECU. The purpose is to convert the objective brightness value of each microLED pixel into a quantitative score of how dazzling the human eye feels, that is, based on the glare evaluation standard of the International Commission on Illumination CIE, the subjective glare value TI caused by a single pixel in the pixelized car light to the human eye is calculated: , Where, is the equivalent screen brightness, which is formed by the superposition of the scattered light of the glare source; is the background adaptive brightness, and q is the angle between the glare source and the line of sight.

[0035] In addition, according to the time of day, it is divided into night, dawn / sunset two cases: At night, directly use to calculate, the constant is based on a large number of human factor engineering experiments and visual psychophysical model fitting of the CIE (International Commission on Illumination) organization; 65 is the scale constant, which is used to map the formula calculation result to the glare index range that meets the subjective evaluation of the human body, which comes from early experiments (such as Schmidt-Clausen and Bindels, 1974) and subsequent CIE technical reports on the statistical fitting of glare tolerance of motor vehicle drivers; 0.25 is the contribution weight of the equivalent screen brightness to the effective adaptive brightness, which is commonly used in early glare models such as CIE 31-1976 and CIE 146-2002, and is also used in 150:2017; 0.8 is a non-linear perception correction index, which reflects the non-linear perception characteristics of the human eye to brightness contrast. At dusk / dawn, add 0.1 rad scattering correction to theta to avoid underestimating glare, the calculation formula is: .

[0036] Then, the subjective glare value distribution map obtained by calculation is smoothed using a Gaussian kernel to generate a smooth weight map, the purpose is to eliminate quantization jaggies and high-frequency noise caused by camera jitter, so that the subsequent optimization algorithm faces a soft and coherent sensitivity cloud map, rather than a noisy point full of burrs.

[0037] Finally, process the glare sensitivity distribution map to generate a smooth weight map for dimming control, that is, a 256x64 smooth weight map, which acts as a penalty term in the following step four, the system projects light to the high weight area, the penalty is greater, so as to output the optimal solution that illuminates itself and does not glare others.

[0038] It should be noted that there is a complex nonlinear relationship between the subjective perception of human eyes to glare and objective brightness, direct control of brightness cannot guarantee comfort, the single-pixel glare formula in the CIE 150:2017 standard is introduced, which simulates the influence of the equivalent light curtain brightness formed by the scattering of incident light in the eye on visual contrast, and maps the physical brightness of each pixel to a subjective discomfort quantitative index conforming to the physiological and psychophysical model of human vision, as a control basis, so that the dimming strategy is improved from ensuring not to see to ensuring to see comfortably.

[0039] Step three, geometric compensation of the glare sensitivity distribution map: Use a cubic polynomial to fit the lane line and calculate the lane line curvature and slope angle in real time, the purpose is to find that the driver of the oncoming vehicle actually appears more left / right in the curve than in the straight line; The cubic polynomial fitting of the lane line is: , x is the longitudinal distance in the vehicle forward direction, and y is the transverse distance in the vehicle forward direction; Based on the standard definition formula for calculating the curvature of a two-dimensional plane curve in differential geometry, the curvature calculation formula is obtained: , Wherein, is the first derivative, which represents the slope of the lane line, is the second inverse, which represents the bending degree of the lane line; Based on the geometric slope definition and the numerical difference method, the slope angle calculation formula is obtained: , Wherein, is a small forward distance interval; Based on IEEE Trans. ITS 2018 “Real-Time Road Surface Reconstruction for ADAS”, a 3x3 perspective transformation matrix is generated according to curvature and slope, the purpose is to project the straight line hypothesis dark area obtained in step two into the real curve perspective, to ensure that the dark area always covers the eyes of the opposite driver like a shadow, and will not leak light due to the widening of the curve.

[0040] Based on Gonzalez, R. C., & Woods, R. E. (2018). Digital Image Processing. Pearson, bilinear interpolation is used for pixel-level resampling, the purpose is to make the projected dark area edge smooth under the premise of controllable calculation amount, avoid step-like mutation causing secondary glare, specifically: As shown in Figure 2 , first locate four field points:

[0041] Then twice linear difference in X direction: At (top row), according to the pixel values of and , the value of point is calculated by interpolation: , At (bottom row), according to the pixel values of and , the value of point is calculated by interpolation: ; Then, linear interpolation in Y direction: Using the values of R1 and R2, interpolate in the y direction to get the pixel value of the final target point : .

[0042] It should be noted that on the curve and slope, the camera imaging will produce distortion due to perspective projection, resulting in a deviation between the dark area located in the image and the target position on the actual three-dimensional road surface. By constructing a three-dimensional road coordinate system and performing inverse projection, the core principle is to solve the inverse process of the perspective projection equation. By using the known lane line equation and camera parameters, the two-dimensional image points are mapped back to the three-dimensional world coordinates, thereby restoring the accurate position of the target in the real world and eliminating the positioning error caused by the perspective.

[0043] Step four, based on the compensation of the smooth weight map in step three and the road lighting demand of the ego vehicle, a dimming instruction for controlling the brightness of each pixel of the pixelated vehicle lamp is generated through an optimization algorithm, specifically: The smooth weight map and the road lighting demand map of the ego vehicle are constructed as a joint state, which aims to let the reinforcement learning network see both demands at once, rather than only one side. The calculation formula is: ; Wherein, is the sensitivity of the oncoming vehicle to glare, is the road lighting demand of the ego vehicle, R is the risk weighting coefficient, which is a dynamic variable calculated in real time according to the ego vehicle state (such as speed, steering, weather), when the ego vehicle is in a high-risk scene (such as high speed, curve), the lighting demand is amplified through R>1 to tell the network that the priority of ensuring the safety of the ego vehicle is higher at the current moment; 256x64x2 is the joint state constructed according to the pixel resolution of 256x64 and the channel number of 2.

[0044] The joint state is input into an offline trained multi-agent deep deterministic policy gradient MADDPG network, and the network outputs a pixel-level pulse width modulation PWM duty cycle instruction as the dimming instruction; the purpose is to give 256x64 PWM duty cycles within 3ms, so that each pixel of the microLED array can independently decide how bright it is, achieving sub-centimeter precision of light and dark area interlacing, rather than the traditional solution of turning on or off the entire block.

[0045] The inference result is directly written into the register of the LED driving chip, the purpose is to update the light state to the theoretically least dazzling new pattern before the next 50ms period comes, the whole closed-loop delay is not more than 10ms, ensuring that the light pattern update speed matches the relative speed (>120km / h scene), avoiding glare leakage.

[0046] It should be noted that the multi-vehicle meeting scene is a typical multi-agent cooperation problem, and the traditional single-agent reinforcement learning cannot solve the challenge of non-static strategy of the other party, while the MADDPG algorithm adopts a centralized training and distributed execution framework, that is, during training, the Critic network can obtain the observation and action information of all vehicles, thereby learning the collaborative strategy; during execution, each Actor network can make optimal decisions only by relying on its own observation. In this embodiment, the joint state is input into the Actor network, which fundamentally avoids game conflicts.

[0047] Step five, execute the dimming instruction in step four to suppress the glare of the oncoming vehicle driver.

[0048] Step six, degradation mode, specifically: Continuously monitor the V2X heartbeat. If no CAM message is received from the other vehicle within 200 ms, immediately set the communication loss flag. This is to let the system know that collaborative optimization is no longer feasible and it must revert to single-vehicle mode. Next, call the preset MobileNetV3-Small lightweight network to quickly estimate an alternative glare sensitivity distribution map using the camera ROI area. This is to ensure that even if V2X data is lost, the system can still roughly know which pixel position the other driver might be sitting in, ensuring that the function is not interrupted. Immediately afterward, display an "ADB degrade prompt" on the instrument panel and record the fault code. The mode switching strategy should refer to the vehicle system functional safety standard "ISO 26262". This is to remind the driver that the system has entered single-vehicle vision fallback mode for easy after-sales traceability. At the same time, continue to execute steps three to five.

[0049] Example 2 like Figure 3 As shown, the dynamic glare suppression system based on multi-vehicle collaborative perception includes a V2X communication module, a processing unit, and a pixelated headlight driving module. The V2X communication module receives perception information including a glare sensitivity distribution map sent by oncoming vehicles. The processing unit runs a human visual sensitivity model, processes the glare sensitivity distribution map to generate a smooth weight map, and runs an optimization algorithm to generate pixel-level dimming instructions based on the smooth weight map and the vehicle's road lighting requirements. It also includes performing dynamic perspective transformation to correct the glare sensitivity distribution map based on lane lines and slope information obtained from the vehicle's sensors, and, when the V2X communication module is interrupted, calling a pre-set lightweight neural network to generate an alternative pixelated glare sensitivity distribution map. The headlight driving module receives the dimming instructions and drives the pixelated headlights to perform corresponding brightness changes.

[0050] Example 3 A vehicle equipped with a dynamic glare suppression system based on multi-vehicle cooperative perception, as described in Embodiment 2.

[0051] Example 4 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the dynamic glare suppression method based on multi-vehicle cooperative perception as described in Embodiment 1.

[0052] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A dynamic glare suppression method based on multi-vehicle cooperative perception, characterized in that, The method comprises the following steps: S1, acquiring multi-source perception information, wherein at least the perception information sent by at least one oncoming vehicle through vehicle wireless communication V2X is acquired; the perception information sent by the oncoming vehicle comprises a glare sensitivity distribution map based on the visual angle of the object vehicle; S2, constructing a human eye glare sensitivity model, processing the glare sensitivity distribution map, and generating a smooth weight map for dimming control; S3, performing geometric compensation on the glare sensitivity distribution map according to a curve or slope; S4, based on the smooth weight map compensated in S3 and the road lighting demand of the ego vehicle, generating dimming instructions for controlling the brightness of each pixel of the pixelated vehicle lamp through an optimization algorithm; S5, executing the dimming instructions in S4 to suppress glare for the oncoming vehicle driver.

2. The method of claim 1, wherein the method is based on multi-vehicle cooperative perception. The step S1 is specifically: S11, real-time pulling of the longitudinal speed, yaw rate, steering wheel angle and turn signal state of the ego vehicle through the CAN bus to construct an ego vehicle motion trajectory prediction model; S12, receiving the CAM message sent by the oncoming vehicle through the C-V2X vehicle-mounted unit, analyzing the 256*64 grayscale map carried therein, and obtaining the glare sensitivity distribution map; When there are multiple object vehicles, different fusion weights are assigned to the glare sensitivity distribution maps sent for each oncoming vehicle according to the distance between the ego vehicle and each oncoming vehicle , and based on the fusion weights , the multiple glare sensitivity distribution maps are linearly superimposed to generate a unified total sensitivity distribution map : ， wherein, , is the longitudinal distance between the host vehicle and the kth oncoming vehicle; is the most appropriate pixel gray map given for each oncoming vehicle. S13, the ego vehicle front camera continuously outputs the oncoming vehicle contour, lane line and road surface slope angle, and constructs a three-dimensional road surface coordinate system according to the lane line cubic polynomial and slope; S14, the millimeter wave radar synchronously gives the distance and speed redundancy, and uses the cascade matching and dynamic weight fusion algorithm to do cross verification with the radar point cloud, to ensure that the sensitive grayscale map corresponds to the real oncoming vehicle.

3. The method of claim 2, wherein the method is based on multi-vehicle cooperative perception. The S13 is specifically: S131, converting the image coordinates into road surface local coordinates, and the calculation formula is: ， Wherein, (u, v) is the pixel coordinate of the center of the opposite vehicle bounding box in the image, K is the camera intrinsic matrix, and d is the actual distance corresponding to the pixel. , ) is the transformed local coordinate of the road surface. S132, compensating the slope and curvature, and the calculation formula is: ， wherein, is the height of the camera installation, and a is the road slope angle. S133, mapping the two-dimensional sensitivity map to a three-dimensional road surface, the formula being: S133, mapping the two-dimensional sensitivity map to a three-dimensional road surface, the formula being: 。 4. The method of claim 1, wherein the method further comprises: The human eye visual sensitivity model in S2 is specifically: S21, based on the glare evaluation standard of the International Commission on Illumination CIE, calculating the subjective glare value TI caused by a single pixel in the pixelated vehicle lamp to the human eye: ， where, is the equivalent veiling luminance, is the background adaptation luminance, θ is the angle between the glare source and the line of sight; 65 is a scale constant, 0.25 is the equivalent veiling luminance is the contribution weight to the effective adaptation luminance, 0.8 is a non-linear perception correction exponent; S22, using Gaussian kernel to smooth the subjective glare value distribution map obtained by calculation to generate a smooth weight map.

5. The method of claim 3, wherein the method further comprises: The compensation processing of S3 is specifically: S31, fitting the lane line with a cubic polynomial and calculating the lane line curvature and slope angle in real time: S311, the lane line is fitted by a cubic polynomial as: , x is a longitudinal distance in the vehicle advancing direction, and y is a transverse distance in the vehicle advancing direction. S312, based on the standard definition formula of two-dimensional plane curve curvature calculated by differential geometry, the curvature calculation formula is obtained: ， wherein, is a first derivative, representing the slope of the lane line, is a second inverse, representing the degree of curvature of the lane line; S313, based on the geometric slope definition and the numerical difference method, the slope angle calculation formula is obtained: ， wherein is a small forward distance interval; S32, generating a 3*3 perspective transformation matrix according to the curvature and the slope; S33, using bilinear interpolation to do pixel-level resampling, specifically: S331, positioning four field points: S332, twice linear difference in X direction: At place, the value of the point is calculated by interpolation according to the pixel values of and . ， At place, the value of the point is calculated by interpolation according to the pixel values of and . ； S333, linear interpolation in Y direction: Using the values of R1 and R2, interpolation is performed in the y direction to obtain the pixel value of the final target point ​ 。 6. The method of claim 1, wherein: The generation of dimming instructions in S4 through an optimization algorithm is specifically: The smooth weight map and the road lighting demand map of the ego vehicle are constructed into a joint state, and the calculation formula is: ； wherein, is the sensitivity to glare from oncoming vehicles, is the self-vehicle road lighting demand, R is a risk weighting factor, and 256x64x2 is a joint state constructed with a pixel resolution of 256x64 and a channel number of 2. The joint state is input into a multi-agent deep deterministic policy gradient MADDPG network trained offline, and the pixel-level pulse width modulation PWM duty cycle instruction output by the network is used as the dimming instruction.

7. The method of claim 1, wherein: Further comprising a degradation step S6, specifically: Real-time monitoring of the V2X communication link state; When V2X communication is interrupted, a lightweight neural network based visual perception model is enabled to estimate an alternative glare sensitivity map from images captured by the ego vehicle camera; Based on the alternative glare sensitivity map, steps S2 to S5 are performed.

8. A suppression system based on the dynamic glare suppression method based on multi-vehicle cooperative perception according to any one of claims 1 to 7, characterized in that, Comprising: a V2X communication module for receiving perception information containing a glare sensitivity map sent by a leading vehicle; a processing unit for: a. running a human eye visual sensitivity model to process the glare sensitivity map to generate a smooth weight map; b. running an optimization algorithm to generate pixel-level dimming instructions based on the smooth weight map and the ego vehicle's road lighting needs; c. performing a dynamic perspective transformation to correct the glare sensitivity map based on lane line and slope information obtained from the ego vehicle's sensors; d. invoking a pre-stored lightweight neural network to generate an alternative glare sensitivity map when the V2X communication module is interrupted; a pixelated vehicle light driving module for receiving the dimming instructions and driving the pixelated vehicle light to perform corresponding brightness changes.

9. A vehicle characterized by comprising: equipped with a multi-vehicle cooperative perception based dynamic glare suppression system as claimed in claim 8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps of the multi-vehicle cooperative perception based dynamic glare suppression method as claimed in any one of claims 1 to 7.

Images