Intelligent vehicle light target detection method, device and storage medium

Abstract

The present application discloses a method, device and storage medium for intelligent headlight target detection, and relates to the technical field of headlight adjustment and target detection. The method includes constructing a low-light image data set; constructing a target detection network model; using the low-light image data set to train the target detection network model to obtain a trained target detection network model; based on the visual perception system of the intelligent vehicle, real-time acquisition of the low-light image to be detected; inputting the low-light image to be detected into the trained target detection network model to obtain the confidence result corresponding to the low-light image to be detected; adjusting the light level of the intelligent vehicle according to the confidence result corresponding to the low-light image to be detected. The present application can not only accurately perceive the target in front of the vehicle in a low-light environment, but also adjust the level of the headlight in real time and accurately.

Description

Technical Field

[0001] The present application relates to the technical field of vehicle light adjustment and target detection, and in particular to an intelligent vehicle light target detection method, device and storage medium. Background Art

[0002] In low-light environments, images captured by image sensors often suffer from problems such as insufficient brightness, reduced contrast, and missing details. These problems seriously affect the detection and recognition capabilities of the on-board visual perception system for key targets such as vehicles, pedestrians, and lane lines. Since traditional target detection algorithms are mostly trained on high-quality images under normal lighting conditions, their model parameters and feature extraction mechanisms are often difficult to adapt to the degradation characteristics of low-light images. In the face of low-light images with insufficient brightness and significant noise, the detection accuracy of existing algorithms may drop significantly, especially for targets with low contrast, small size, or partially occluded, the risk of detection failure is higher. In addition, low-light environments pose a more severe challenge to the real-time performance of detection algorithms. Due to poor image quality, the algorithm may need to introduce pre-processing modules such as image enhancement and multi-frame fusion, which not only increases the computational complexity but also slows down the detection speed. Therefore, it is of great significance to achieve high-precision, high-robustness, and low-latency target detection in low-light environments.

[0003] The progress of intelligent headlight systems has opened up new possibilities for target detection tasks in low-light environments. The system can adaptively adjust the lighting area, light type and brightness, which not only improves the brightness of the road ahead and the driver's field of vision, but also significantly improves the imaging quality in low-light environments, providing a more favorable working environment for visual perception algorithms. However, the perception ability of existing intelligent headlight systems in dealing with complex low-light scenes is still limited, and they are not yet able to accurately obtain target information and fail to fully meet the needs of target detection. It is urgent to build a collaborative perception system that integrates "intelligent headlight systems" and "visual perception" to achieve a positive feedback loop from "intelligent lighting" to "high-quality images" to "precise perception", thereby improving the vehicle perception accuracy in low-light environments.

[0004] However, there is no precedent in the current related art for combining the headlight system with vehicle perception. In the related art, there is a projection control method for an intelligent headlight system. The method mainly focuses on how to use the vehicle's visual perception system to control the projection content and method of the headlight. The time delay from the perception end to the headlight end is a key factor affecting the performance of the intelligent headlight system. There may be a problem of too long a delay, which causes the headlight projection to be unable to respond to environmental changes in a timely manner. Therefore, the real-time performance of this method is poor and the headlight position cannot be adjusted in real time. In the related art, there is also an intelligent headlight control system, which aims to improve driving safety through automation and reduce traffic accidents caused by driver delays or forgetting to use the headlights. The system integrates antenna modules, visual modules, body control modules, lighting modules, and voice modules to achieve real-time monitoring and intelligent response to the vehicle's surroundings. The system consists of multiple modules, which increases the complexity of the vehicle, which may make the system more difficult to maintain and more difficult to diagnose problems when a fault occurs. At the same time, the system performance is highly dependent on the accuracy of the visual module and the antenna module. If there are errors in the recognition or reception of signals by these modules, false alarms or missed alarms may occur, thereby affecting the reliability of the system. Therefore, the system cannot guarantee a high level of perception accuracy, and cannot accurately adjust the light position.

[0005] Based on this, how to design an intelligent headlight system that can not only accurately sense the target in front of the vehicle in a low-light environment, but also adjust the gear position of the headlights in real time and accurately has become a technical problem that needs to be solved urgently in this field. This is of great significance for promoting the development of intelligent driving technology and enhancing driving safety in low-light environments. Summary of the invention

[0006] The purpose of this application is to provide an intelligent headlight target detection method, device and storage medium, which can not only accurately perceive the target in front of the vehicle in a low-light environment, but also adjust the gear position of the headlights in real time and accurately.

[0007] To achieve the above objectives, this application provides the following solutions.

[0008] In a first aspect, the present application provides an intelligent vehicle light target detection method, and the intelligent vehicle light target detection method comprises the following steps.

[0009] Construct a low-light image dataset; the low-light image dataset includes a plurality of low-light images, and the low-light images are images taken in front of a vehicle in a simulated nighttime driving environment.

[0010] Construct a target detection network model; the target detection network model is a detection model that takes the low-light image as input and outputs a confidence result, the target detection network model includes a feature extraction module, a feature fusion module, an attention mechanism module and a low-light enhancement module, wherein the low-light enhancement module includes a degradation repairer and an illumination estimator, the illumination estimator is used to brighten the low-light image to obtain a brightened image, the degradation repairer is used to obtain an enhanced image based on the brightened image, and the confidence result is a confidence value for identifying pedestrians and / or vehicles from the low-light image.

[0011] The target detection network model is trained using the low-light image dataset to obtain a trained target detection network model.

[0012] Based on the visual perception system of the smart car, the low-light image to be detected is collected in real time; the low-light image to be detected is the image in front of the vehicle collected by the visual perception system at the current light level after the smart car is driving at night and turning on the lights. The visual perception system is a system equipped by the smart car itself for collecting the image in front of the vehicle in real time.

[0013] The low-light image to be detected is input into the trained target detection network model to obtain the confidence result corresponding to the low-light image to be detected.

[0014] The light level of the smart car is adjusted according to the confidence result corresponding to the low-light image to be detected.

[0015] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described intelligent vehicle light target detection methods.

[0016] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described intelligent vehicle light target detection methods.

[0017] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0018] The present application provides a method, device and storage medium for intelligent headlight target detection. By constructing a target detection network model composed of a feature extraction module, a feature fusion module, an attention mechanism module and a dark light enhancement module, not only a target detection network model based on an attention mechanism is implemented, but also a dark light enhancement module composed of a degradation repairer and an illumination estimator is set in the target detection network model. By combining the illumination estimator with the degradation repairer to perform low-light enhancement on low-light images, and matching the attention mechanism module, the accuracy of target detection in low-light environments is effectively improved. In addition, the present application also collects low-light images to be detected in real time through the visual perception system of the intelligent vehicle, and inputs the low-light images to be detected into the trained target detection network model, so as to obtain the confidence result corresponding to the low-light images to be detected, and accurately identify in real time whether there are pedestrians and / or vehicles in front of the intelligent vehicle, thereby improving the perception accuracy and real-time performance of the intelligent vehicle. Moreover, the light level of the smart car can be further adjusted in real time according to the confidence result corresponding to the low-light image to be detected. Compared with the traditional way in which the driver completely relies on subjective consciousness to adjust the light level, the present application is more accurate and reliable in adjusting the light level according to the confidence result of the low-light image to be detected, thereby achieving the goal of accurately perceiving the target in front of the vehicle in a low-light environment and adjusting the light level in real time and accurately, thus ensuring driving safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 A flowchart of an intelligent vehicle light target detection method provided in accordance with an embodiment of the present application.

[0021] Figure 2 This is a principle flow chart of the active dimming strategy provided in one embodiment of the present application.

[0022] Figure 3 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0024] like Figure 1 As shown, this embodiment provides an intelligent vehicle light target detection method, and the intelligent vehicle light target detection method includes the following steps.

[0025] Step S1: construct a low-light image dataset.

[0026] In this embodiment, the low-light image data set includes a plurality of low-light images, where the low-light images are images taken in front of the vehicle in a simulated nighttime driving environment.

[0027] In this embodiment, step S1 constructs a low-light image data set, which specifically includes the following steps.

[0028] Step S11, using a Vignette filter, performing a low-light image feature restoration process on each of the low-light images in the Nuscence image data set to obtain a plurality of low-light images after restoration.

[0029] Step S12: performing a superimposition process on each of the restored low-light images with a real dark image to obtain a plurality of darkened low-light images.

[0030] Step S13: adding Gaussian noise and Poisson noise to each of the darkened low-light images to obtain a low-light image data set.

[0031] Step S14: Based on the real low-light images in the real nighttime driving environment, each low-light image in the low-light image dataset is evaluated using grayscale, brightness, contrast and saturation as evaluation indicators.

[0032] Step S14 specifically includes the following steps.

[0033] Step S141, setting the grayscale threshold, brightness threshold, contrast threshold and saturation threshold.

[0034] Step S142: Compare the grayscale, brightness, contrast and saturation of each low-light image in the low-light image data set with the real low-light image in the real night driving environment to obtain the grayscale difference, brightness difference, contrast difference and saturation difference between the low-light image and the real low-light image.

[0035] Step S143, respectively determine whether the grayscale difference, the brightness difference, the contrast difference and the saturation difference are less than the grayscale threshold, the brightness threshold, the contrast threshold and the saturation threshold. If so, execute steps S2 and S3. Otherwise, return to step S11 and reconstruct the low-light image data set that meets the evaluation index.

[0036] Step S2: construct a target detection network model.

[0037] In this embodiment, the target detection network model is a detection model that takes a low-light image as input and outputs a confidence result. The target detection network model includes a feature extraction module, a feature fusion module, an attention mechanism module, and a low-light enhancement module, wherein the low-light enhancement module includes a degradation repairer and an illumination estimator, the illumination estimator is used to brighten the low-light image to obtain a brightened image, the degradation repairer is used to obtain an enhanced image based on the brightened image, and the confidence result is a confidence value for identifying pedestrians and / or vehicles from the low-light image.

[0038] In this embodiment, the illumination estimator is used to brighten the low-light image to obtain a brightened image, which specifically includes the following steps: splicing and fusing the low-light image and the illumination prior map corresponding to the low-light image according to the channel dimension to obtain illumination features; generating a brightened map based on the illumination features; and setting the brightened map as a three-channel RGB tensor to generate a brightened image.

[0039] In this embodiment, the degradation repairer is used to obtain an enhanced image based on the brightened image, which specifically includes the following steps: generating a plurality of hierarchical features based on the brightened image; generating a residual image based on the plurality of hierarchical features; and adding the residual image and the brightened image to obtain an enhanced image.

[0040] Step S3: Using the low-light image data set, the target detection network model is trained to obtain a trained target detection network model.

[0041] Step S4: Based on the visual perception system of the smart car, the low-light image to be detected is collected in real time.

[0042] In this embodiment, the low-light image to be detected is an image in front of the vehicle captured by the visual perception system at the current light level after the smart car is driving at night with the lights turned on. The visual perception system refers to a system installed in the smart car itself for capturing images in front of the vehicle in real time.

[0043] Step S5: input the low-light image to be detected into the trained target detection network model to obtain the confidence result corresponding to the low-light image to be detected.

[0044] Step S6: adjusting the light level of the smart car according to the confidence result corresponding to the low-light image to be detected.

[0045] In this embodiment, step S6 adjusts the light level of the smart car according to the confidence result corresponding to the low-light image to be detected, and specifically includes the following steps.

[0046] Step S61: According to the confidence result corresponding to the low-light image to be detected, an active dimming strategy is adopted to adjust the light level of the smart car.

[0047] Figure 2 The principle flow chart of the active dimming strategy in this embodiment is shown. First, the light shifting rules of the smart car are set to include the following: (1) After the smart car turns on the lights, the initial light gear is low beam gear 1. (2) Low beam gear 1 is the lowest gear of low beam, and the light brightness increases from the lowest gear to the highest gear. (3) When the highest gear of low beam still cannot identify pedestrians and / or vehicles from the low-light image, it switches to high beam gear 1. (4) High beam gear 1 is the lowest gear of high beam, and the light brightness corresponding to the lowest gear of high beam is greater than the light brightness corresponding to the highest gear of low beam, and the light brightness increases from the lowest gear to the highest gear. In this embodiment, there is no downshifting. After turning on the lights, it is low beam gear 1, that is, the lowest gear of low beam. If pedestrians and / or vehicles can be detected and identified, the low beam gear 1 will be maintained. Downshifting will result in dimming of the light, and downshifting will become meaningless.

[0048] like Figure 2 As shown, in this embodiment, step S61 adopts an active dimming strategy to adjust the light level of the smart car according to the confidence result corresponding to the low-light image to be detected, and specifically includes the following steps.

[0049] Step S611: compare the confidence value of the to-be-detected low-light image under the current light level with the confidence threshold, determine whether the confidence value of the to-be-detected low-light image is greater than the confidence threshold, and obtain a comparison result.

[0050] When the comparison result is yes, the current light level is kept unchanged.

[0051] When the comparison result is no, the current light gear is upgraded to the next gear, and the process returns to step S4 to collect the low-light image to be detected after the light gear is adjusted to the next gear, and input the image into the trained target detection network model to obtain the corresponding confidence value, and then determine the relationship between the confidence value and the confidence threshold until the confidence value of the low-light image to be detected is greater than the confidence threshold.

[0052] This embodiment focuses on target detection tasks in low-light environments. Based on the intelligent vehicle lighting system platform, it can achieve high-precision and high-robustness target detection in low-light environments, and adjust the vehicle lighting position in real time and accurately to improve driving safety.

[0053] In order to make the technical solution of the present application clearer, the specific implementation process of the technical solution of the present application is explained below in the form of examples, including the following specific implementation steps.

[0054] Step 1: Build a low-light image dataset.

[0055] This embodiment constructs a low-light image dataset, generates an image in front of a vehicle in a low-light environment by darkening the image and simulating noise addition, and evaluates the dataset. Specifically, the following steps are included.

[0056] S1-1: Using the NuScence image dataset as the basis, it is split into training set, validation set, and test set in a ratio of 7:2:1.

[0057] S1-2: Use the Vignette filter of Open CV (open source computer vision library) to restore the characteristics of low-light images.

[0058] The Vignette filter uses an X / Y bidirectional one-dimensional Gaussian distribution function to assign a greater weight to pixels closer to the center and a smaller weight to pixels farther away, thereby achieving a gradual decrease in brightness and saturation from the center to the periphery. This processing method can better simulate the light distribution in a real low-light environment, with the brightness gradually decaying from the center of the image to the edge, achieving a smoother brightness decay while retaining the details of the central area of ​​the image.

[0059] S1-3: Superimpose the real dark image to reduce the overall image brightness. Further darkening is performed by superimposing the real dark image that meets the low-light characteristics. The details are shown in formula (1).

[0060]

[0061] in, This is the image after superimposing the real dark image. To meet the real dark image of low light characteristics, is the original image, α is the image weight, the larger the weight is, the closer the image is to the original image brightness.

[0062] S1-4: For real pictures in a low-light environment, Gaussian noise and Poisson noise often exist, so this embodiment needs to add simulated noise. First, Gaussian noise is added, and its probability density function can be expressed as formula (2).

[0063]

[0064] in, is the probability density function of adding Gaussian noise; x is a random variable, i.e., the actual value of the noise; μis the mean of the Gaussian distribution, that is, the average value of the noise. In many cases, the mean is 0, that is, the noise is zero-mean Gaussian noise. σ is the standard deviation of the Gaussian distribution, which represents the degree of dispersion of the noise. The larger the standard deviation, the greater the amplitude of the noise. is the variance of the Gaussian distribution, which represents the degree of noise dispersion; e is the base of natural logarithms.

[0065] Gaussian noise will be generated when there is no stable light source in the image shooting scene, resulting in uneven brightness, the mutual influence of various circuit components in the imaging system, and the image sensor working for a long time, resulting in excessive temperature and affecting the sensing performance. Then add Poisson noise, and its probability density function can be expressed as formula (3).

[0066]

[0067] in, P ( x = k ) represents the probability density function of adding Poisson noise, k Indicates the number of times an event occurs within a specific time interval or area. k is a non-negative integer; λ is the parameter of Poisson distribution, which indicates the average number of events in a given time interval or region. This parameter is also called the mean and variance of Poisson distribution; ! is the factorial symbol.

[0068] S1-5: Evaluation of low-light image datasets.

[0069] This embodiment selects grayscale, brightness, contrast and saturation of basic image attributes as evaluation indicators, and compares the self-made low-light image dataset with the real low-light image dataset in the real night driving environment. By comparing the grayscale, brightness, contrast and saturation of the self-made low-light image dataset with the real low-light image dataset in the real night driving environment, it is evaluated and judged whether the self-made low-light image dataset meets the conditions of the night driving environment. If it does not meet the conditions of the night driving environment, repeat step one to supplement the dataset, thereby ensuring the accuracy and reliability of data processing in subsequent steps.

[0070] In view of the fact that existing open source low-light image datasets are less focused on autonomous driving scenarios, and the self-collected datasets are costly, difficult, and of low quality, this embodiment constructs an artificial low-light image dataset based on image darkening and simulated noise addition, uses Vignette filter and superimposes real dark images to reduce image brightness, and simulates noise on the darkened images, thereby constructing a low-light image dataset. Finally, the low-light image dataset is evaluated to provide a reliable data basis for subsequent algorithm design.

[0071] Step 2: Design an object detection network based on the attention mechanism.

[0072] In order to solve the problems of poor quality of low-light images and difficulty in feature extraction, and to avoid the cumbersome multi-stage training process, this embodiment constructs a one-stage retinex-based framework (ORF), which consists of an illumination estimator and a degradation repairer. The original image enhancement model is modified by introducing perturbation terms in the reflectivity and illumination to suppress noise, artifacts, underexposure, overexposure, color distortion and other problems, and can model image defects well.

[0073] S2-1: Lighting Estimator First, the image S and S Illuminance prior map Concatenate by channel dimension and then fuse through 1 conv1×1 (one-dimensional convolution). Because well-exposed areas can provide semantic context information for underexposed areas, in order to generate illumination features ,image S and illumination prior map The splicing is fused through conv1×1 and then through a depth-separable conv5×5 to simulate the interaction between regions under different lighting conditions. Then, the lighting estimator Use 1 conv1×1 to fuse the illumination features , to generate a brightened image , will brighten the image Set it to a three-channel RGB tensor to improve its representation ability in simulating nonlinearity between RGB channels for color enhancement. Used to brighten images S , generating a brightened image .

[0074] S2-2: Design an IGT (Illumination-Guided Transformer) module. In order to model long-distance dependencies, this embodiment proposes an IGT module as a degradation repairer. The basic unit of the degradation repairer is IGAB (Illumination-Guided Attention Block), which consists of two layer normalizations LN (Layer Normalization), an IG-MSA (Illumination-Guided Multi-head Self-Attention) and a feed-forward network FFN (Feed-Forward Network). The IGA (Illumination-Guided Attention) module effectively utilizes the illumination information by integrating the IG-MSA mechanism, and incorporates the illumination information into the calculation of the attention weight, so that the model can better handle low-light areas and enhance the feature expression ability of the model.

[0075] In this embodiment, the degradation repairer adopts a three-scale U-shaped architecture, and its input is the brightened image . Introducing hierarchical features , expressed as formula (4).

[0076]

[0077] in, represents the hierarchical features, H Indicates the height of the image. W Indicates the width of the image. Ch Indicates the distance from the center point of the image to the edge or axis.

[0078] In the downsampling branch, the brightened image After 1 conv3×3, the initial hierarchical features are generated , initial stratification features After 1 IGAB and 1 conv4×4 with a stride of 2 (for downsampling features), the first layer features are generated , the first layer feature After 2 IGABs and 1 conv4×4 with a step size of 2, the second layer feature is generated , the second layer feature After two IGABs, a structure symmetrical to the downsampling branch is designed as the upsampling branch, and deconv2×2 with a step size of 2 is used to upsample features. Skip connections are also used in the network structure of the IGT module to reduce the information loss caused by the downsampling branch. The upsampling branch outputs the residual image , residual image And the input brightened image Add together to get the final enhanced image , expressed as formula (5).

[0079]

[0080] in, represents the final enhanced image, represents the residual image, Represents the brightened image.

[0081] S2-3: Insert the degradation repairer into the illumination estimator to form a dark light enhancement module. This process is a plug-and-play network module, which can be understood as a series process. The main function is to better process low-light areas through the dark light enhancement module, enhance image quality, and make image processing and analysis in low-light environments more accurate and reliable.

[0082] S2-4: Design a feature extraction module and a feature fusion module, which specifically includes the following steps.

[0083] S2-4-1: Hierarchical contextualization: The input feature map undergoes linear projection and a series of deep convolutional layers to generate a multi-level feature map, each layer capturing contextual information at different scales.

[0084] S2-4-2: Gated aggregation: Gated weights are generated through linear layers, and features at different levels are aggregated by element-wise weighted summation, that is, contextual information is selectively aggregated into the modulator of each query token.

[0085] S2-4-3: Perform element-wise affine transformation: Inject the aggregated modulator into each query token through affine transformation.

[0086] In the traditional Focal Nets model, the modulator of the query token is used to adjust and enhance the influence of different levels of contextual information in the feature map on each query token. The modulator is a mechanism for controlling the weights of different contextual information in the feature map. In the Focal Nets model, it is generated through a gated aggregation layer to perform weighted aggregation of multi-level features. When generating the modulator, the Focal Nets model first generates a multi-level feature map through linear projection and deep convolution layers. Each layer of feature map captures contextual information at different scales. Then, the Focal Nets model generates a modulator through a gated aggregation layer. The modulator weights the features at different levels through gated weights so that each query token can be adjusted according to the specific contextual information. When injecting the modulator, the aggregated modulator is injected into each query token through an element-by-element affine transformation. This process is accomplished through an affine transformation (such as a linear transformation and a bias term). In this way, the role of the modulator is to adjust the features of the query token so that it can better capture the contextual information, thereby improving the performance of the model. The modulator of the query token is actually injected into the query token when it is generated, with the aim of optimizing the feature representation of the query token by adjusting and enhancing the contextual information in the feature map. The modulator helps the model selectively fuse information in different contexts and improves its responsiveness to various scales and contextual information.

[0087] S2-5: Design the attention mechanism module, which specifically includes the following steps.

[0088] S2-5-1: Multi-scale parallel subnetwork: EMA (Enhanced Multi-scale Attention) module is a module used to capture features of different scales in images and enhance the model's feature expression capabilities. The EMA module includes input feature separation, multi-scale feature extraction, feature fusion, and enhanced multi-scale attention mechanism. The EMA module effectively improves the model's feature expression capabilities in processing complex scenes by combining information of different scales and attention mechanisms.

[0089] In this embodiment, the EMA module uses two parallel sub-networks, using 1×1 convolution kernel and 3×3 convolution kernel respectively to process the input features, where the 1×1 convolution kernel is used for local interactions within the channel and the 3×3 convolution kernel is used to capture broader spatial relationships.

[0090] S2-5-2: Feature reshaping and grouping: Reshape some channel dimensions into batch dimensions and group the channel dimensions into multiple sub-features to ensure that each group of features is evenly distributed in space, improve the diversity and discriminability of feature representation, and avoid some form of dimensionality reduction through general convolution.

[0091] S2-5-3: Cross-space learning: Use 2D global average pooling on the output of the 1×1 branch to obtain the global spatial feature descriptor, and then perform a matrix dot product operation on the global spatial feature descriptor and the output of the 3×3 branch to calculate the similarity between each pixel and the global spatial feature, thereby highlighting pixels related to the global features and suppressing irrelevant pixels.

[0092] This embodiment adopts a dark light enhancement module that combines an illumination estimator with a degradation repairer to perform low-light enhancement on low-light images, adopts a feature extraction module and a feature fusion module to extract and fuse features, and uses an attention mechanism module to improve detection accuracy and robustness.

[0093] Step 3: Train the target detection network model to obtain target detection weights for low-light environments.

[0094] In this embodiment, a self-made low-light image dataset is trained on the NVIDIA GeForce RTX3090 graphics card and the Intel(R) Core(TM) i9-10980XE CPU@3.00GHz ×36 platform, and then the attention-based target detection network model designed in step 2 is used to perform 500 rounds of training with the help of the Ubuntu operating system. After the training, the optimal parameters of the target detection network model are obtained to obtain the target detection weight in the low-light environment.

[0095] Step 4: Design an active dimming strategy based on the intelligent vehicle lighting system.

[0096] The existing Adaptive Front-lighting System (AFS) only considers the adaptive adjustment of the headlight illumination distance to the vehicle speed. However, in real driving scenarios under low-light conditions, changes in the headlight illumination intensity and distance will affect the detection of the target by the visual perception system. At the same time, if the headlight is adjusted to an overly strong position at a too close distance, it may cause a safety hazard of dazzling other people on the road. The active dimming strategy based on the intelligent headlight system designed in this embodiment is as follows: Figure 2 As shown, the specific steps include the following steps.

[0097] S4-1: First, determine whether the confidence value output at the current moment is greater than the set confidence threshold. In order to avoid too frequent adjustment of the headlights, the output confidence value is smoothed and fitted; different Cof _ exp The threshold affects the selection of gear. Cof _ expThe higher the confidence threshold, the more accurate and longer the distance the target can be detected, making driving safer. However, the overall headlight level will be higher, which may cause the risk of glare to road traffic participants. Cof _ exp They are set to 0.2, 0.4, and 0.6 respectively. Cof _ exp Indicates the set confidence threshold.

[0098] S4-2: The main tasks of driving a vehicle are different in different relative distance segments. In this embodiment, 0-37.5m is set as the close distance segment. In the close distance segment, pedestrian detection needs to be considered. In this stage, the confidence of detecting pedestrians and vehicles needs to be comprehensively considered, and the smaller value is taken as Cof _ out In this embodiment, 35-100m is set as the long-distance segment. In the long-distance segment, only vehicle detection needs to be considered. Cof _ out is the confidence value of the detected pedestrians and / or vehicles.

[0099] Step 5: Design the intelligent lighting system.

[0100] The intelligent headlight system used in this embodiment includes a position sensor, an ambient light sensor, a control unit, a headlight module, a power module, and a communication module. Among them, the position sensor, the ambient light sensor, the headlight module, the power module, and the communication module are connected to the control unit respectively. The ambient light sensor is used to detect the light intensity of the external environment and transmit the detected light intensity data to the control unit. The control unit includes a high-performance microcontroller, such as the STM32 (embedded single-chip microcomputer) series. The control unit is used to receive the light intensity data of the external environment detected by the ambient light sensor, and run an adaptive light adjustment algorithm to dynamically adjust the light output according to the real-time lighting conditions. The headlight module includes a headlight, which includes a low beam and a high beam. The low beam and the high beam can be high-brightness LED (Light Emitting Diode) lamps or laser lamps, and can adjust the brightness and direction of the low beam and the high beam through a stepper motor according to the instructions of the control unit. The headlight is then fed back to the control unit through a position sensor. The power module provides a stable power supply for each component of the system, and the power module has the functions of efficient energy conversion, low power consumption, and stable output. The communication module is used for data transmission and communication interaction, mainly to feed back system status and adjustment information to the driver.

[0101] In this embodiment, the intelligent vehicle lighting system includes the following specific contents.

[0102] S5-1: Detect the external light intensity in real time through the ambient light sensor and transmit the data to the control unit.

[0103] S5-2: The control unit processes the ambient light data according to the preset light adjustment algorithm involved in this embodiment and based on the external light intensity detected by the ambient light sensor to calculate the required light output.

[0104] S5-3: The control unit sends instructions to the headlight module according to the calculated required light output, and controls the headlight module to adjust the brightness and direction of the headlight.

[0105] S5-4: While adjusting the brightness and direction of the lights, the system status and adjustment information are fed back to the driver through the communication module, and manual adjustment and monitoring can be performed through the vehicle information system or smartphone.

[0106] S5-5: The power module provides stable power to all parts of the system to ensure the normal operation of the system.

[0107] S5-6: Integrate the above-mentioned headlight components into an intelligent headlight module to form an intelligent headlight system.

[0108] This embodiment significantly improves the safety and comfort of driving in low-light environments by detecting the ambient light intensity in real time and automatically adjusting the brightness and direction of the headlights. While ensuring the target detection effect and safe braking distance in low-light environments, the brightness of the headlights is maintained at the minimum illumination level, minimizing interference to other vehicles and pedestrians.

[0109] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store intelligent headlight target detection data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for intelligent headlight target detection is implemented.

[0110] Those skilled in the art will understand that Figure 3The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0111] In an exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps of the intelligent vehicle light target detection method when executing the computer program.

[0112] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the steps of the intelligent vehicle light target detection method when executed by a processor.

[0113] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0114] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. An intelligent vehicle light target detection method, characterized in that: The intelligent vehicle light target detection method comprises: Constructing a low-light image dataset; the low-light image dataset includes a plurality of low-light images, and the low-light images are images taken in front of a vehicle in a simulated nighttime driving environment; Constructing a target detection network model; the target detection network model is a detection model that takes the low-light image as input and outputs a confidence result, the target detection network model includes a feature extraction module, a feature fusion module, an attention mechanism module and a low-light enhancement module, wherein the low-light enhancement module includes a degradation repairer and an illumination estimator, the illumination estimator is used to brighten the low-light image to obtain a brightened image, the degradation repairer is used to obtain an enhanced image based on the brightened image, and the confidence result is a confidence value of identifying pedestrians and / or vehicles from the low-light image; Using the low-light image data set, training the target detection network model to obtain a trained target detection network model; Based on the visual perception system of the smart car, the low-light image to be detected is collected in real time; the low-light image to be detected is the image in front of the vehicle collected by the visual perception system at the current light level after the smart car is driving at night and turning on the lights. The visual perception system is a system installed by the smart car itself for collecting the image in front of the vehicle in real time; Inputting the low-light image to be detected into the trained target detection network model to obtain the confidence result corresponding to the low-light image to be detected; According to the confidence result corresponding to the low-light image to be detected, adjusting the light level of the smart car; According to the confidence result corresponding to the low-light image to be detected, adjusting the light level of the smart car specifically includes: According to the confidence result corresponding to the low-light image to be detected, an active dimming strategy is adopted to adjust the light level of the smart car; Setting the light shifting rules of the smart car includes: After the smart car turns on the lights, the initial light gear is low beam first gear; The first low beam gear is the minimum low beam gear, and the light brightness increases from the minimum low beam gear to the maximum low beam gear; When the maximum low beam gear still cannot identify pedestrians and / or vehicles from the low-light image, switch to the first high beam gear; The high beam first gear is the minimum high beam gear, the light brightness corresponding to the minimum high beam gear is greater than the light brightness corresponding to the maximum low beam gear, and the light brightness increases from the minimum high beam gear to the maximum high beam gear; According to the confidence result corresponding to the low-light image to be detected, an active dimming strategy is adopted to adjust the light level of the smart car, specifically including: Comparing the confidence value of the to-be-detected low-light image under the current light level with the confidence threshold, determining whether the confidence value of the to-be-detected low-light image is greater than the confidence threshold, and obtaining a comparison result; When the comparison result is yes, keep the current light level unchanged; When the comparison result is no, the current light gear is upgraded to the next gear, and the process returns to the step of "real-time acquisition of low-light images to be detected based on the visual perception system of the smart car" until the confidence value of the low-light image to be detected is greater than the confidence threshold.

2. The intelligent vehicle light target detection method according to claim 1, characterized in that: Construct a low-light image dataset, including: Using a Vignette filter, performing low-light image feature restoration processing on each of the low-light images in the Nuscence image dataset to obtain a number of low-light images after restoration processing; Performing a superimposition process on each of the restored low-light images to obtain a plurality of darkened low-light images; Gaussian noise and Poisson noise are added to each of the darkened low-light images to obtain a low-light image data set.

3. The intelligent vehicle light target detection method according to claim 2, characterized in that: After the step of adding Gaussian noise and Poisson noise to each of the dimmed low-light images to obtain a low-light image data set, the intelligent vehicle light target detection method further includes: According to real low-light images in a real nighttime driving environment, each low-light image in the low-light image dataset is evaluated using grayscale, brightness, contrast and saturation as evaluation indicators.

4. The intelligent vehicle light target detection method according to claim 1, characterized in that: The illumination estimator is used to brighten the low-light image to obtain the brightened image, which specifically includes: The low-light image and the illumination prior map corresponding to the low-light image are spliced ​​and merged according to the channel dimension to obtain illumination features; Generate a brightening image according to the illumination characteristics; The brightened image is set as a three-channel RGB tensor to generate a brightened image.

5. The intelligent vehicle light target detection method according to claim 1, characterized in that: The degradation repairer is used to obtain an enhanced image according to the brightened image, and specifically includes: Generating a plurality of hierarchical features according to the brightened image; generating a residual image according to the plurality of hierarchical features; The residual image and the brightened image are added to obtain the enhanced image.

6. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent vehicle light target detection method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent vehicle light target detection method according to any one of claims 1 to 5 is implemented.

Images