Intelligent driving car headlight performance dynamic test method, equipment and storage medium

Abstract

The present invention provides a method, device and storage medium for dynamic testing of the performance of headlights of intelligent driving vehicles, including the following steps: S1, obtaining dynamic test data and performing data processing and feature extraction; S2, building a test model by combining a hybrid neural network model structure with a feature pyramid network; S3, training and debugging the model of step S2 using the data processed by step S1; S4, finally, performing performance evaluation and verification on the model trained in step S3 and using it to dynamically test the performance of headlights of intelligent driving vehicles. In the field of intelligent driving vehicles, the present invention introduces a dynamic testing system and method for headlight performance, which significantly improves the fit and comprehensiveness of test results with actual driving scenarios compared to traditional static tests. This innovation not only profoundly affects the safety and user experience of intelligent driving vehicles, but also promotes the intelligentization and standardization of lighting technology in the entire automotive industry.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile safety testing, and in particular to a method, equipment and storage medium for dynamic testing of headlamp performance of an intelligent driving automobile. Background Art

[0002] With the rapid development of intelligent driving vehicles, traditional static testing methods for automobile headlight performance can no longer meet the increasingly complex and changing driving needs. Static testing methods mainly focus on evaluating the performance of headlights when the vehicle is stationary, and test the brightness, distribution and compliance of the headlights through a fixed test environment and preset lighting conditions. However, this testing method ignores the complex dynamic environment that intelligent driving vehicles face during actual driving, such as speed changes, road bumps, cornering and interactions with other traffic participants, all of which will significantly affect the lighting effect and safety of the headlights.

[0003] The core of smart driving cars lies in the high degree of interaction between the vehicle and the surrounding environment. This feature requires that the headlight system must not only have basic lighting functions, but also be able to make intelligent adjustments based on real-time road conditions and driving needs to ensure driving safety. However, the existing static test methods are powerless when faced with these intelligent functions. On the one hand, the standards of static tests are often too idealistic and fail to fully consider the complexity and variability in actual driving, resulting in a large deviation between the test results and the actual usage scenarios; on the other hand, with the continuous innovation of headlight technology, such as the application of new light sources such as LED and laser, and the popularization of intelligent technologies such as adaptive front lighting systems (AFS) and matrix LED headlights, the update speed of existing detection technology has lagged far behind the development of headlight technology, and it is impossible to conduct a comprehensive and accurate performance evaluation of new headlights. Summary of the invention

[0004] The present invention provides a method, device and storage medium for dynamically testing the performance of headlights of an intelligent driving vehicle, which can solve at least one of the technical problems in the background technology.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for dynamically testing the performance of a headlamp of an intelligent driving vehicle comprises the following steps:

[0007] S1. Obtain dynamic test data and perform data processing and feature extraction;

[0008] S2. Construct a test model by combining the hybrid neural network model structure with the feature pyramid network;

[0009] S3, using the data processed in step S1 to train and debug the model in step S2;

[0010] S4. Finally, the performance of the model trained in step S3 is evaluated and verified and used to dynamically test the performance of the headlights of the intelligent driving car.

[0011] Furthermore, step S1 specifically includes:

[0012] S11. Data cleaning: First, the collected raw data is cleaned to remove outliers or noise data;

[0013] S12. Data standardization: Standardize the data to eliminate the influence of external factors, so that the data measured at different times and in different environments are comparable. The Min-Max standardization method is used to obtain standardized data:

[0014] Min-Max Normalization:

[0015] is the standardized data value, X is the original data value, is the minimum value in the data set, is the maximum value in the data set.

[0016] S13. Feature extraction: Extract key features from the cleaned and standardized data, including the maximum, minimum, average, and standard deviation of light intensity, as well as the trend of light intensity changes over time, vehicle speed, and target approach angle. These features will be used as input to the neural network model to evaluate the performance of the headlights.

[0017] Furthermore, the model construction in step S2 includes adopting a hybrid neural network model structure to integrate the deep separable convolution and feature pyramid network into the convolutional neural network, specifically including:

[0018] S21, depth convolution means applying a convolution kernel to each input channel independently without mixing channel information;

[0019] For the input feature map ,high H ,width W , number of channels , deep convolution application convolution kernels, each convolution kernel only acts on the corresponding input channel;

[0020] Output feature map Each channel c Calculated by the following formula:

[0021]

[0022] in k is the convolution kernel size,s is the step length; It acts on c The convolution kernel of the input channels, σ is the activation function, Represents the input feature map c Channels at position ( x , y ) value;

[0023] S22, point-by-point convolution uses a 1x1 convolution kernel to mix the output of the depth convolution to generate a new feature map;

[0024] Point-by-point convolution uses a 1×1 convolution kernel to mix the output channels of the depth convolution; the output feature map Each position ( x , y , c ') is calculated by the following formula:

[0025]

[0026] in, is the weight matrix of point-wise convolution, is the depth convolution output feature map at position ( x , y ),aisle c The value of .

[0027] Furthermore, the model construction in step S2 also includes using a feature pyramid network FPN, introducing a feature pyramid network FPN structure in the deep layer of the convolutional neural network CNN, fusing high-level semantic information with low-level detail information through top-down paths and lateral connections, and generating a multi-scale feature map, as follows:

[0028] S231, bottom-up path: This is the forward propagation process of the convolutional neural network. Each layer generates feature maps of different scales. These feature maps gradually reduce in size as the depth of the network increases, but the semantic information is richer.

[0029] S232, top-down path: through upsampling, Upsampled to The same spatial dimensions give ; The spatial resolution of high-level feature maps is enlarged to match the size of feature maps of lower layers;

[0030] S233, Horizontal connection: The upsampled high-level feature map is merged with the low-level feature map of the corresponding size by element-by-element addition or other fusion methods; this merging not only retains the low-level detail information, but also incorporates the high-level semantic information to generate an enhanced multi-scale feature map; that is, the horizontal connection specifically and The fused feature map is obtained by fusion in the set way ;

[0031] S234, Feature representation: Through the integration process, a series of fused multi-scale feature maps are obtained, which are represented as , ,…, , where N is the number of layers involved in the fusion in FPN; each They all contain rich information from high-level to low-level, including both fine spatial details at the low level and abstract semantic information at the high level;

[0032] These feature maps are represented as in are the height, width and number of channels of the feature map respectively.

[0033] Furthermore, the model construction in step S2 also includes adding a spatial attention module before the feature map output by the feature pyramid network FPN enters the recurrent neural network RNN. The spatial attention module generates a weight value for each spatial position by analyzing the feature map to dynamically adjust the importance of the input feature.

[0034] Furthermore, the model construction in step S2 also includes using a long short-term memory network LSTM or a gated recurrent unit GRU to capture time dependency.

[0035] Furthermore, the model construction in step S2 also includes fusing the spatial multi-scale features obtained by combining the convolutional neural network CNN with the feature pyramid network FPN with the temporal features of the recurrent neural network RNN ​​enhanced by the spatial attention module;

[0036] Through concatenation or weighted sum, these two parts of features are effectively combined to form a comprehensive description of the headlight performance.

[0037] Furthermore, in step S3, during the training process, the gradient of the loss function with respect to the model parameters is calculated by the back propagation algorithm, and the parameter values ​​are updated accordingly.

[0038] Loss Function: Mean Squared Error (MSE) ;

[0039] in, The true value is the actual measured light intensity. The predicted value is the light intensity after being predicted by the model. is the sample size;

[0040] Optimization algorithm: Gradient descent method uses Adam update rule:

[0041] in, θ are model parameters, η is the learning rate, and It is a modified first- and second-order moment meter.

[0042] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0043] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0044] It can be seen from the above technical solutions that the intelligent driving car headlamp performance dynamic test method of the present invention introduces a dynamic headlamp performance test system and method in the field of intelligent driving cars, which significantly improves the fit and comprehensiveness of the test results with the actual driving situation compared to the traditional static test. This innovation not only has a profound impact on the safety and user experience of intelligent driving cars, but also promotes the intelligentization and standardization of lighting technology in the entire automotive industry.

[0045] The intelligent driving car of the present invention integrates high-precision sensors and real-time data analysis technology, and its headlamp performance dynamic test system can simulate and adapt to various complex and changeable road and traffic environments. This includes but is not limited to scenarios such as lateral approach on the driver and passenger side under dynamic road conditions, and rapid longitudinal approach of target vehicles or pedestrians, and achieves instant adjustment of lighting brightness, range and angle. This innovation not only ensures driving safety at night and in low-light environments, but also reduces the interference of glare to other road users through intelligent optimization of lighting strategies, significantly improving the overall safety of road driving.

[0046] The intelligent driving car headlamp performance dynamic test system of the present invention combines multi-dimensional information such as vehicle speed, road conditions, weather conditions, etc. to conduct a comprehensive and in-depth performance evaluation of the headlamp. The system can calculate and feedback the driver's average recognizable distance and safe reaction time in real time, providing a more accurate decision-making basis for the intelligent driving assistance system. At the same time, through continuous data collection and analysis, the system can continuously optimize the lighting strategy, realize personalized lighting experience, and meet the needs and preferences of different drivers.

[0047] The development and application of the intelligent driving car headlight performance dynamic test system of the present invention provides strong support for the intelligent transformation of automotive lighting technology. It not only promotes the innovation and development of lighting technology, but also promotes the formulation and improvement of relevant international standards. Through standardized testing processes and evaluation indicators, the performance of intelligent driving car headlights can be uniformly recognized and standardized worldwide, providing a solid foundation for cross-brand and cross-platform intelligent interconnection. In addition, the widespread application of this system will also accelerate the transformation and upgrading of the automotive lighting industry and promote the entire industry to develop in a more efficient, safe and environmentally friendly direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flowchart of an embodiment of the present invention;

[0049] Figure 2 This is a data cleaning flow chart of an embodiment of the present invention;

[0050] Figure 3 This is a data standardization flow chart of an embodiment of the present invention;

[0051] Figure 4 A flow chart is constructed for the model of the embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of a real-life laboratory of an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of a subjective evaluation test scenario according to an embodiment of the present invention;

[0054] Figure 7 It is a schematic diagram of the arrangement of dummy sensors under the longitudinal approach working condition of an embodiment of the present invention;

[0055] Figure 8 It is a schematic diagram of the arrangement of dummy sensors under the lateral approach condition according to an embodiment of the present invention;

[0056] Fig. 9 A schematic diagram of an intelligent mobile platform according to an embodiment of the present invention;

[0057] Fig.10 Schematic diagram of the overall experimental process of an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0059] like Figure 1As shown, the intelligent driving vehicle headlamp performance dynamic test method described in this embodiment specifically includes the following steps:

[0060] S1, obtain test data and perform data processing and feature extraction;

[0061] S2. Construct a test model by combining the hybrid neural network model structure with the feature pyramid network;

[0062] S3, using the data processed in step S1 to train and debug the model in step S2;

[0063] S4. Finally, the performance of the model trained in step S3 is evaluated and verified and used to dynamically test the performance of the headlights of the intelligent driving car.

[0064] The following are detailed descriptions:

[0065] 1. Data Processing

[0066] During the dynamic test of the headlight performance of intelligent driving vehicles, high-precision light intensity sensors are placed on the simulated dummy to simulate the lighting conditions in real driving situations. These data cover the light intensity information under different speeds, different target approach angles, different road conditions and traffic scenes, providing a rich data basis for subsequent performance evaluation.

[0067] Data preprocessing:

[0068] 1. Data cleaning: such as Figure 2 As shown, first, the collected raw data is cleaned to remove abnormal values ​​or noise data caused by sensor failure, environmental interference, etc., to ensure the accuracy and reliability of the data. (Identify abnormal values ​​by setting thresholds or statistical methods)

[0069] 2. Data standardization: Since light intensity data may be affected by many factors (such as time, weather, sensor sensitivity, etc.), it is necessary to standardize the data to eliminate the influence of these factors so that the data measured at different times and in different environments are comparable. The Min-Max standardization method is used to obtain standardized data. Figure 3 As shown;

[0070] Min-Max Normalization:

[0071] is the standardized data value, X is the original data value, is the minimum value in the data set, is the maximum value in the data set.

[0072] 3. Feature extraction: Extract key features from the cleaned and standardized data, including the maximum, minimum, average, and standard deviation of light intensity, as well as the trend of light intensity changes over time, vehicle speed, target approach angle, etc. These features will be used as input to the neural network model to evaluate the performance of the headlights.

[0073] 2. Neural Network Construction

[0074] 1. Neural network model construction:

[0075] like Figure 4 As shown in the figure, in order to more accurately evaluate the performance of smart driving car headlights in different dynamic scenarios, a neural network model is used to conduct in-depth analysis and evaluation of the pre-processed light intensity data. Neural networks have unique advantages in complex system performance evaluation due to their powerful nonlinear fitting and self-learning capabilities.

[0076] (1) Light intensity data conversion

[0077] 2D image construction: First, the light intensity data measured by the sensor at multiple locations on the dummy is organized into a series of 2D images. Each image represents the light intensity distribution at a point in time, where the grayscale value or color intensity of each pixel is directly mapped to the light intensity at that location. This approach allows the use of CNN's powerful spatial feature extraction capabilities.

[0078] Temporal sequence alignment: Ensure that all image sequences are aligned in time so that the RNN can correctly capture temporal dependencies. This may require interpolation or resampling of the original data.

[0079] (2) Standardization and normalization

[0080] Normalizing image data to ensure that pixel values ​​are distributed within a reasonable range (such as 0 to 1) helps speed up the training process and improve model performance. Normalizing time series data makes data at different time points or between different sensors comparable.

[0081] When evaluating the performance of headlights, a hybrid neural network model was constructed that combines the advantages of convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The CNN part effectively extracts the spatial distribution characteristics of the headlight illumination intensity, such as edges, textures, etc., through multi-layer convolution and pooling operations. The RNN part is responsible for capturing the changes in these illumination features over time, including changes in vehicle speed, target approach angles, or dynamic adjustments to the illumination pattern caused by obstacles. The feature fusion layer combines the spatial features of CNN with the temporal features of RNN to ensure that the model can fully understand the spatial distribution and temporal variation characteristics of the headlight lighting effect. Finally, the output layer generates an evaluation result based on the fused features through a fully connected layer. The evaluation result is a lighting effect score, thereby achieving an accurate evaluation of the performance of the headlights.

[0082] Partial model training: The preprocessed light intensity data is used as the training set to train the neural network model. During the training process, the model parameters (such as weights, biases, etc.) are continuously adjusted so that the model can accurately predict the lighting effect of the headlights in different scenarios. At the same time, cross-validation and other methods are used to evaluate the generalization ability of the model to ensure that its performance on unknown data is stable and reliable. The preprocessed light intensity data is divided into training set, validation set and test set. The training set is used to train the model, the validation set is used to adjust the model parameters to prevent overfitting, and the test set is used to evaluate the final performance of the model. Then, a suitable loss function is selected to guide the model training process. For the headlight performance evaluation task, the mean square error (MSE) regression loss function is selected to measure the difference between the light intensity predicted by the model and the actual measured value. The gradient descent method (Adam) is used to optimize the model parameters. During the training process, the gradient of the loss function with respect to the model parameters is calculated by the back-propagation algorithm, and the parameter values ​​are updated accordingly.

[0083] Loss Function: Mean Squared Error (MSE) ;

[0084] in, The true value is the actual measured light intensity. The predicted value is the light intensity after being predicted by the model. is the sample size;

[0085] Optimization algorithm: Gradient descent method such as Adam update rule (taking Adam as an example):

[0086] in, θ are model parameters, η is the learning rate, and It is a modified first- and second-order moment meter.

[0087] Specifically, in the evaluation of the performance of smart driving car headlights, combining the hybrid neural network model structure with the feature pyramid network (FPN) can significantly improve the model's ability to evaluate the performance of headlights under complex lighting conditions. The following is a model construction scheme that combines these technical details with the evaluation of headlight performance. Specifically, a hybrid neural network model structure is used to integrate the deep separable convolution and feature pyramid network into the convolutional neural network, including:

[0088] 1. CNN part is integrated with FPN:

[0089] 1.1 CNN initial layer: Traditional convolutional layers in the initial layer often consume a lot of computing resources and parameters because they learn spatial features and inter-channel correlations at the same time. Depthwise separable convolution reduces these costs through two steps of depthwise convolution and pointwise convolution.

[0090] 1.1.1 Depth Convolution: Apply a convolution kernel to each input channel independently without mixing channel information. This reduces the number of parameters and computation because each channel is processed by only one convolution kernel.

[0091] For the input feature map (high H ,width W , number of channels ), deep convolution application convolution kernels, each of which acts only on the corresponding input channel.

[0092] Output feature map (generally ,in p is filling, k is the convolution kernel size, s is the step size) for each channel c It can be calculated by the following formula:

[0093]

[0094] in, It acts on c The convolution kernel of the input channels, σ is the activation function (such as ReLU), Represents the input feature map c Channels at position ( x , y ). Note that this simplifies boundary handling (such as padding), and in practice you may need to adjust the index to account for padding.

[0095] 1.1.2 Point-by-point convolution: Use a 1x1 convolution kernel to mix the output of the depthwise convolution to generate a new feature map. This step is responsible for learning the correlation between channels.

[0096] Point-wise convolution uses a 1×1 convolution kernel to mix the output channels of the depthwise convolution. Output feature map Each position ( x , y , c ′) can be calculated by the following formula:

[0097]

[0098] in, is the weight matrix of point-wise convolution, is the depth convolution output feature map at position ( x , y ),aisle c The value of .

[0099] 1.1.3 Advantages of depthwise separable convolution: Depthwise separable convolution is used in the initial layer to replace the traditional convolution layer. The combination of depthwise convolution and point-by-point convolution can effectively reduce the amount of calculation and parameters, while maintaining a strong ability to extract spatial features of light intensity.

[0100] 1.2 CNN deep part: Using feature pyramid network (FPN), the FPN structure is introduced in the deep layer of CNN, and high-level semantic information is fused with low-level detail information through top-down paths and lateral connections to generate multi-scale feature maps. These multi-scale feature maps can capture the changes in light intensity at different distances and regions, which is crucial for evaluating the lighting range and effect of headlights.

[0101] 1.3FPN part

[0102] 1.3.1FPN is implemented through the following steps:

[0103] ①. Bottom-up path: In the forward propagation process of CNN, each layer generates feature maps of different scales. These feature maps gradually reduce in size as the network depth increases (through pooling or convolution with a step size greater than 1), but the semantic information is richer.

[0104] ②Top-down path: Through upsampling, Upsampled to The same spatial dimensions give The spatial resolution of the high-level feature map is enlarged to match the size of the feature map of the lower layer. The purpose of this path is to pass the semantic information of the high-level to the low-level.

[0105] ③ Horizontal connection: The upsampled high-level feature map is merged with the low-level feature map of the corresponding size by element-by-element addition (or other fusion methods). This merging not only retains the low-level detail information, but also incorporates the high-level semantic information to generate an enhanced multi-scale feature map. and By fusing in some way (such as element-by-element addition or concatenation and convolution), the fused feature map is obtained. .

[0106] ④ Feature representation: Each fused feature map contains rich multi-scale information, including both low-level fine spatial details and high-level abstract semantic information. These feature maps are crucial for evaluating the lighting range and effect of headlights because they can capture changes in light intensity at different distances and regions. For example, low-level feature maps may be better at identifying the edges and textures of nearby objects, while high-level feature maps can better understand the layout and lighting distribution of the entire scene. Through the above-mentioned FPN integration process, the present invention obtains a series of fused multi-scale feature maps, which can be represented as , ,…, , where N is the number of layers involved in the fusion in FPN. They all contain rich information from high-level to low-level, including both fine spatial details at the low level and abstract semantic information at the high level.

[0107] The exact dimensions of these feature maps may vary depending on the network design and input size, but can generally be expressed as in are the height, width and number of channels of the feature map respectively.

[0108] 2. Attention Layer (spatial attention module): Before the feature map output by FPN enters the RNN, a spatial attention module is added. This module generates a weight value for each spatial position by analyzing the feature map to dynamically adjust the importance of the input feature. In this way, the model can focus more on key areas in the image (such as areas with significant changes in light intensity), thereby improving the accuracy of the evaluation.

[0109] 3.RNN part: Use long short-term memory network (LSTM) or gated recurrent unit (GRU) to capture time dependencies. Considering that the performance of headlights may be affected by multiple time-varying factors such as vehicle speed and road conditions, LSTM or GRU can effectively learn and predict these dynamic changes.

[0110] In order to further improve the performance, a bidirectional LSTM or bidirectional GRU can be used to consider the context information at the same time, so as to have a more comprehensive understanding of the change pattern of headlamp performance in time series.

[0111] 4. Feature fusion layer: The spatial multi-scale features after combining CNN and FPN are fused with the RNN temporal features enhanced by the spatial attention module. Through concatenation or weighted sum, these two parts of features are effectively combined to form a comprehensive description of the headlight performance.

[0112] Specific training steps: In intelligent systems, continuous learning and incremental learning are the key to enabling models to dynamically adapt to new data and environmental changes. In view of the headlamp performance evaluation scenario, combined with the principle of incremental learning, the following specific implementation steps are designed:

[0113] Dynamic data loading implementation steps

[0114] ① Design data flow:

[0115] Data interface design: Establish a standardized data interface that can receive new data from sensors, logging systems, or other data sources in real time.

[0116] Data storage mechanism: Set up data cache or message queue to buffer data arriving in real time and avoid system overload.

[0117] Data trigger mechanism: Define data trigger thresholds to trigger the data loading process when the amount of new data reaches a certain scale or the time interval meets the conditions.

[0118] ②Update the dataset:

[0119] Data cleaning and preprocessing: Perform necessary cleaning and preprocessing on newly received data to ensure the consistency and accuracy of the data format.

[0120] Data integration: Integrate the cleaned new data with the existing data set. You can consider timestamps or serial numbers as the basis for data sorting and integration.

[0121] Data consistency check: During the integration process, a data consistency check is performed to ensure that there are no logical conflicts between the new and old data.

[0122] ③Analyze the characteristics of new data:

[0123] Data distribution analysis: Perform statistical analysis on new data to understand its data distribution and characteristics, including but not limited to mean, standard deviation, outliers, etc.

[0124] Feature extraction: Extract key features relevant to new data and evaluate the impact of these features on headlamp performance.

[0125] ④Determine the training goal:

[0126] Task definition: Based on the characteristics of the new data, clarify the specific goals that the training model needs to achieve, such as improving recognition accuracy under specific lighting conditions.

[0127] Evaluation indicator setting: Set reasonable evaluation indicators for the training process, such as accuracy, recall rate, F1 score, etc., to monitor the training effect.

[0128] ⑤Adjust model structure or parameters:

[0129] Model structure optimization: Based on the characteristics of the new data and the training objectives, consider whether the model structure needs to be adjusted, such as adding convolutional layers, changing activation functions, etc.

[0130] Parameter adjustment: Adjust the initial parameters of the model, such as learning rate, batch size, etc., according to the data characteristics to better adapt to the training of new data.

[0131] 4. Dynamic performance evaluation:

[0132] 1. Lighting effect evaluation: Use the trained neural network model to evaluate new light intensity data and predict the lighting effect of headlights in different scenarios.

[0133] 2. Continuous optimization and improvement: Based on the evaluation results of the neural network model, the headlamp system is continuously optimized and improved. By adjusting the design parameters of the headlamp and optimizing the control algorithm, its performance and safety in actual use can be further improved. At the same time, new test data can also be included in the training set to continuously update and improve the neural network model to adapt to the rapid development of the headlight evaluation system.

[0134] The following are examples:

[0135] First, a laboratory was built to simulate the real road conditions of night driving. The length × width × height (m) of the laboratory is 150 × 39 × 7 (clearance), and an asphalt road surface that simulates the real situation is used, such as Figure 5 As shown in the figure, a lifting optical test screen is set up at 50m, and a natural light source simulation system is used to achieve stepless adjustment of the illumination in the experimental area. Figure 6 As shown, in conjunction with the customized development of the target, an advanced target movement system is used, such as Fig. 9 A dummy model equipped with an illuminance sensor is specially designed, as shown in Figure 7 , Figure 8 As shown, in order to more accurately simulate and test the visual recognition situation in the night driving environment.

[0136] In the test scenario, a series of subjective evaluation test conditions are defined, including:

[0137] 1. Lateral approach on the driver's side: Set the target speed to 6.5km / h and the simulated vehicle speeds to 20 / 40 / 60 / 80km / h respectively.

[0138] 2. Lateral approach on the passenger side: Set the target speed to 5km / h and the simulated vehicle speed to 20 / 40 / 60 / 80km / h.

[0139] 3. Target object approaches longitudinally: simulate relative speed in the range of 0 to 80 km / h.

[0140] In order to comprehensively test the performance of illumination sensors at different parts of the dummy model, the present invention makes the following arrangements:

[0141] Under the longitudinal approach condition, the illumination sensors of the dummy model are respectively arranged on the head, chest, abdomen, left upper arm, left forearm, right upper arm, right forearm, left thigh, left calf, right thigh and right calf.

[0142] Under the lateral approach condition, the sensor layout is slightly simplified, but still covers key parts such as the head, left upper arm, left forearm, right upper arm, right forearm, left thigh, left calf, right thigh, and right calf.

[0143] At the same time, an intelligent mobile platform was designed and developed, which has a high-precision positioning function, can realize the movement of the dummy according to the planned path, and realize spatial positioning in the digital laboratory. The platform also has the power supply capability for data acquisition and sensor equipment, supports remote control of multiple signal source inputs, and greatly improves the automation and intelligence level of the test.

[0144] In the specific test process, the speed of the target object and the vehicle speed under three working conditions were recorded respectively to obtain the relative speed between the two. Then, the light recognizable distance was tested four times, recorded as a, b, c, d, and the average recognizable distance [(a+b+c+d) / 4] was calculated. Finally, the formula:

[0145] Safety reaction time = average recognizable distance / relative speed, to calculate the safety reaction time, which is a key assessment indicator in the vision safety evaluation system.

[0146] The specific calculation process is as follows: different working conditions are designed, which are divided into three working conditions. Working condition one is that the target object approaches from the driving side, the target object speed is 6.5km / h, the vehicle speeds are 20km / h, 40km / h, 60km / h, 80km / h, and the relative speeds at this time are 21km / h, 40.5km / h, 60.4km / h, 80.3km / h, respectively. In these cases, their recognizable distance 1, recognizable distance 2, recognizable distance 3, and recognizable distance 4 are measured respectively, and then the average recognizable distance is calculated, and finally the safe reaction time (average recognizable distance / relative speed) is calculated.

[0147] Working condition 2 is that the target object approaches longitudinally, the target object speed is 0km / h, the vehicle speeds are 20km / h, 40km / h, 60km / h, 80km / h respectively, and the relative speeds at this time are 20km / h, 40km / h, 60km / h, 80km / h respectively. In these cases, their recognizable distance 1, recognizable distance 2, recognizable distance 3, and recognizable distance 4 are measured respectively, and then the average recognizable distance is calculated, and finally the safe reaction time (average recognizable distance / relative speed) is calculated.

[0148] Working condition three is that the target object approaches from the passenger side, the target object speed is 5km / h, the vehicle speeds are 20km / h, 40km / h, 60km / h, 80km / h respectively, and the relative speeds at this time are 20.6km / h, 40.3km / h, 60.2km / h, 80.2km / h respectively. In these cases, their recognizable distance 1, recognizable distance 2, recognizable distance 3, recognizable distance 4 are measured respectively, and then the average recognizable distance is calculated, and finally the safe reaction time (average recognizable distance / relative speed) is calculated.

[0149] The overall experimental design process is as follows Fig.10 shown.

[0150] The vehicle under test should be equipped with headlights that meet the national mandatory requirements. Before the test, the front and rear axle loads of the vehicle need to be measured and the total vehicle mass needs to be calculated, and this weight is regarded as the curb weight of the vehicle. The test equipment is installed on the passenger side of the vehicle. After the test equipment is installed, it should be loaded and the requirements of the following formula must be met:

[0151] M+G+T+P=(M+200)×(100%±1%)

[0152] Among them, M is the curb weight, G is the driver's weight, T is the test equipment weight, and P is the load weight, all in kilograms. At the same time, in order to minimize the impact of the vehicle's body height drop after the equipment is installed, the distribution of the load should be adjusted so that the height difference of the fenders of the four wheels of the vehicle is kept within 0.003m when the driver is included; if this requirement cannot be met after the load adjustment, the difference between the load rate of the front and rear axles of the vehicle and the load rate when full of fuel and no load is ensured to be less than 1%.

[0153] In order to meet the accuracy requirements, make the test data reflect the actual lighting level of the tested vehicle headlights as realistically as possible, improve the accuracy of the test data and ensure the repeatability of the test, the vehicle and equipment must meet the following requirements during the test:

[0154] 1. After the test starts, the speed of the vehicle under test remains stable, and the speed deviation is within the range of ±2km / h during the entire test process;

[0155] 2. Adjust the steering wheel so that the longitudinal center line of the vehicle under test always remains in the middle of the test lane, and the lateral distance between the center point of the front edge of the vehicle and the center line of the lane remains within the range of ±0.1m.

[0156] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0157] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0158] In another embodiment provided in the present application, a computer program product including instructions is also provided. When the computer program product is run on a computer, the computer executes any one of the methods for dynamically testing the performance of headlights of an intelligent driving vehicle in the above embodiments.

[0159] It is understandable that the system, device and storage medium provided in the embodiments of the present invention correspond to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above methods.

[0160] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0161] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0162] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0163] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic test method for the performance of headlights of an intelligent driving vehicle, characterized in that: The following steps are included: S1. Obtain dynamic test data and perform data processing and feature extraction; S2. Construct a test model by combining the hybrid neural network model structure with the feature pyramid network; S3, using the data processed in step S1 to train and debug the model in step S2; S4, finally, the performance evaluation and verification of the model trained in step S3 is performed and used to dynamically test the performance of the headlights of the intelligent driving car; The dynamic test data in S1 includes information on light intensity at different speeds, different target approach angles, different road conditions, and traffic scenarios; The model construction in step S2 includes adopting a hybrid neural network model structure to integrate the deep separable convolution and feature pyramid network into the convolutional neural network, specifically including: S21, depth convolution means applying a convolution kernel to each input channel independently without mixing channel information; For the input feature map ,high H ,width W , number of channels , deep convolution application convolution kernels, each of which acts only on the corresponding input channel; Output feature map Each channel c Calculated by the following formula: Where k is the convolution kernel size; is the convolution kernel acting on the cth input channel, σ is the activation function, Represents the value of the cth channel of the input feature map at position (x, y); S22, point-by-point convolution uses a 1x1 convolution kernel to mix the output of the depth convolution to generate a new feature map; Point-by-point convolution uses a 1×1 convolution kernel to mix the output channels of the depth convolution; the output feature map Each position ( x , y , c ') is calculated by the following formula: in, is the weight matrix of point-wise convolution, is the depth convolution output feature map at position ( x , y ),aisle c The value of The model construction in step S2 also includes the use of feature pyramid network FPN, which introduces the feature pyramid network FPN structure into the deep layer of convolutional neural network CNN, and fuses high-level semantic information with low-level detail information through top-down paths and lateral connections to generate multi-scale feature maps, as follows: S231, bottom-up path: This is the forward propagation process of the convolutional neural network. Each layer generates feature maps of different scales. These feature maps gradually reduce in size as the depth of the network increases, but the semantic information is richer. S232, top-down path: through upsampling, Upsampled to The same spatial dimensions give ; Enlarge the spatial resolution of high-level feature maps to match the size of low-level feature maps; S233, Horizontal connection: The upsampled high-level feature map is merged with the low-level feature map of the corresponding size by element-by-element addition or other fusion methods; this merging not only retains the low-level detail information, but also incorporates the high-level semantic information to generate an enhanced multi-scale feature map; that is, the horizontal connection specifically combines and The fused feature map is obtained by fusion in the set way ; S234, Feature representation: Through the integration process, a series of fused multi-scale feature maps are obtained, which are represented as , ,…, , where N is the number of layers involved in the fusion in FPN; each They all contain rich information from high-level to low-level, including both fine spatial details at the low level and abstract semantic information at the high level; These feature maps are represented as in are the height, width and number of channels of the feature map respectively.

2. The method for dynamic testing of headlight performance of an intelligent driving vehicle according to claim 1, characterized in that: Step S1 specifically includes: S11. Data cleaning: First, the collected raw data is cleaned to remove outliers or noise data; S12. Data standardization: Standardize the data to eliminate the influence of external factors, so that the data measured at different times and in different environments are comparable. The Min-Max standardization method is used to obtain standardized data: Min-Max Normalization: is the standardized data value, X is the original data value, is the minimum value in the data set, is the maximum value in the data set; S13. Feature extraction: Extract key features from the cleaned and standardized data, including the maximum, minimum, average, and standard deviation of light intensity, as well as the trend of light intensity changes over time, vehicle speed, and target approach angle. These features will be used as input to the neural network model to evaluate the performance of the headlights.

3. The method for dynamic testing of headlight performance of an intelligent driving vehicle according to claim 1, characterized in that: The model construction in step S2 also includes adding a spatial attention module before the feature map output by the feature pyramid network FPN enters the recurrent neural network RNN. The spatial attention module generates a weight value for each spatial position by analyzing the feature map to dynamically adjust the importance of the input feature.

4. The method for dynamic testing of headlight performance of an intelligent driving vehicle according to claim 1, characterized in that: The model building in step S2 also includes using a long short-term memory network LSTM or a gated recurrent unit GRU to capture time dependencies.

5. The method for dynamic testing of headlamp performance of an intelligent driving vehicle according to claim 1, characterized in that: The model construction in step S2 also includes fusing the spatial multi-scale features obtained by combining the convolutional neural network CNN with the feature pyramid network FPN with the temporal features of the recurrent neural network RNN ​​enhanced by the spatial attention module; Through concatenation or weighted sum, these two features are effectively combined to form a comprehensive description of the headlamp performance.

6. The method for dynamic testing of headlamp performance of an intelligent driving vehicle according to claim 1, characterized in that: In step S3, during the training process, the gradient of the loss function with respect to the model parameters is calculated by the back propagation algorithm, and the parameter values ​​are updated accordingly. Loss Function: Mean Squared Error ; in, The true value is the actual measured light intensity. The predicted value is the light intensity after being predicted by the model. is the sample size; Optimization algorithm: Gradient descent method uses Adam update rule: in, θ are model parameters, η is the learning rate, and It is a modified first- and second-order moment meter.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 6.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 6.

Images