HDV driving intention recognition method and system based on lamp language under mixed traffic flow

Abstract

The invention belongs to the technical field of automatic driving, and particularly relates to an HDV driving intention recognition method and system based on lamp language under mixed traffic flow, and the method comprises the steps: constructing an automatic driving network which comprises a network-connected automatic driving vehicle (CAV), an artificial driving vehicle (HDV) and an MEC node; the CAV obtains a lamp language image of a surrounding target HDV; transmitting the lamp language image to an MEC node, and preprocessing the lamp language image by the MEC node; performing target HDV driving intention recognition on the preprocessed lamp language image by adopting a trained driving intention recognition model based on the YOLO-v7 network; the HDV driving intention is transmitted to the CAV; the CAV generates a corresponding driving decision according to the HDV driving intention and executes the driving decision; according to the invention, the lamp signal of the HDV is identified in real time by using the YOLO-v7 algorithm, so that the accuracy of driving intention identification is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of automatic driving, and specifically relates to a method and system for recognizing HDV driving intention based on light language in mixed traffic flow. Background Art

[0002] As the urban traffic environment where connected autonomous vehicles (CAVs) and human-driven vehicles (HDVs) coexist becomes more and more popular, how to improve the safety and efficiency of traffic flow has become the focus of current research in the field of transportation. Intelligent connected vehicles can interact with the surrounding environment through vehicle networking technology, but in mixed traffic flow, since HDVs cannot communicate directly with CAVs, CAVs find it difficult to accurately predict and identify the driving intentions of HDVs, thus affecting their decision-making and driving safety.

[0003] Lighting is the main means of non-verbal communication between HDVs and surrounding vehicles, conveying driving intentions through signals such as brake lights and turn signals. However, due to the diversity of light signals and environmental interference, it is difficult for CAVs to accurately identify HDVs' driving intentions through traditional perception methods. Summary of the invention

[0004] In order to solve the problem that it is difficult for CAV to accurately identify the driving intention of HDV in mixed traffic flow in the prior art, the present invention proposes a method for identifying HDV driving intention based on light language in mixed traffic flow, the method comprising: constructing an autonomous driving network, wherein the autonomous driving network comprises a networked autonomous driving vehicle CAV, a manually driven vehicle HDV and a MEC node; the CAV obtains the light language image of the surrounding target HDV; the light language image is transmitted to the MEC node, and the MEC node preprocesses the light language image; the trained driving intention recognition model based on the YOLO-v7 network is used to identify the driving intention of the target HDV on the preprocessed light language image; the HDV driving intention is transmitted to the CAV; the CAV generates a corresponding driving decision according to the HDV driving intention, and executes the driving decision.

[0005] A HDV driving intention recognition system based on light language in mixed traffic flow, the system comprising: a real-time light language image acquisition unit, an HDV driving intention recognition module, a vehicle networking intention sharing service module and a driving behavior decision execution unit;

[0006] The real-time light language image acquisition unit is used to acquire HDV light language images;

[0007] The HDV driving intention recognition module recognizes the driving intention of the HDV based on the collected light language image and the driving intention recognition model based on the YOLO-v7 network on the MEC node;

[0008] The Internet of Vehicles intention sharing service module unicasts the driving intention of the HDV to the target CAV;

[0009] The driving behavior decision execution unit makes decisions and plans paths according to the received HDV driving intention information, and executes corresponding driving behaviors.

[0010] Beneficial effects of the present invention:

[0011] The present invention adopts the YOLO-v7 algorithm to recognize the light language signals of HDV in real time. By introducing an improved feature extraction layer and an optimized network structure, the accuracy of YOLO-v7 in identifying the driving intentions of HDV such as braking, steering, and overtaking in complex traffic environments is improved; by utilizing 5G communication technology, the light language images are transmitted from CAV to MEC nodes in real time for processing, and the identified driving intentions are promptly returned to CAV for decision-making and execution. Compared with traditional cloud computing methods, the use of MEC nodes for distributed computing greatly reduces the delay in data transmission, significantly improves the overall response speed of the system, and reduces the cost of large-scale deployment; by accurately identifying the driving intentions of HDV and sharing them with CAV, CAV can adjust its driving strategy according to real-time information to avoid potential traffic accidents, thereby improving road traffic safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of the HDV driving intention recognition method of the present invention;

[0013] Figure 2 It is a structural schematic diagram of the HDV driving intention recognition system of the present invention;

[0014] Figure 3 It is the technical roadmap of the HDV driving intention recognition system of the present invention;

[0015] Figure 4 It is an overall schematic diagram of the HDV driving intention recognition system of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] A method for identifying HDV driving intention based on light language in mixed traffic flow, the method comprising: constructing an autonomous driving network, wherein the autonomous driving network comprises a connected autonomous driving vehicle CAV, a manually driven vehicle HDV and a MEC node; the CAV obtains light language images of surrounding target HDVs; the light language images are transmitted to the MEC node, and the MEC node preprocesses the light language images; a trained driving intention recognition model based on a YOLO-v7 network is used to identify the target HDV driving intention on the preprocessed light language images; the HDV driving intention is transmitted to the CAV; the CAV generates a corresponding driving decision according to the HDV driving intention, and executes the driving decision.

[0018] The YOLO-v7 algorithm is a target detection algorithm based on a deep convolutional neural network and is widely used in real-time object recognition tasks. YOLO-v7 is known for its single-step detection, high accuracy and speed, and can significantly improve recognition accuracy while maintaining detection speed. The YOLO-v7 algorithm in the present invention further improves the recognition ability of light signals in complex traffic environments by introducing an improved feature extraction layer and an optimized network structure, especially showing excellent performance in multi-target detection and dynamic backgrounds.

[0019] A specific implementation method of a HDV driving intention recognition method based on light language in mixed traffic flow, such as Figure 1 and Figure 2 As shown, specifically including:

[0020] The S1 CAV is equipped with high-resolution cameras and other auxiliary sensors, which are arranged at the front and sides of the vehicle to ensure comprehensive coverage of the environment around the vehicle.

[0021] The camera should have the following features: it can maintain high image quality under complex lighting conditions, such as scenes of transition between strong light and shadow; it can provide a wider viewing angle to ensure that multiple HDV light signals can be captured simultaneously in multi-lane mixed traffic flows; it can still provide clear images at night or in low light conditions, providing high-quality data for subsequent light signal recognition.

[0022] In this embodiment, when the CAV enters a mixed traffic flow area, the on-board visual sensor will begin to capture the light signals of the surrounding HDVs in real time. These light signals include brake lights, turn signals, overtaking lights, etc. The specific acquisition process is: the visual sensor continuously monitors the surrounding environment, identifies and locks the light signal source of the HDV, and ensures that images can be continuously and stably acquired under high-speed driving and complex traffic scenarios; to ensure that the captured images have sufficient timing information, the frame rate of the camera needs to be automatically adjusted according to the actual road conditions to adapt to vehicles of different speeds and environmental changes; before the collected images are transmitted to the MEC node, they are first preliminarily processed in the on-board computing unit. The preliminary processing includes image compression and resolution adjustment to ensure that the image is not distorted during network transmission.

[0023] S2. Transmit the light language image to the MEC node in real time through the 5G communication network, and perform various image preprocessing operations on the MEC node to improve the recognition accuracy of the light language signal. Specifically, CAV transmits the collected light language image to the MEC node in real time through the 5G communication network, ensuring that the light language image data can be quickly and stably transmitted to the edge computing node in a complex traffic environment, reducing transmission delays; combined with error correction technology (forward error correction FEC) and encrypted transmission, the security and integrity of image data are guaranteed.

[0024] In this embodiment, after the MEC node receives the light language image, a preprocessing operation is performed, specifically including: improving the contrast of the image through histogram equalization, and gamma correction is used to adjust the image brightness to make the light language signal clearer and more prominent under different lighting conditions; using the median filtering method to denoise the image, effectively reducing the interference of noise on the recognition of the light language signal, and retaining the edge and detail features of the light language signal; according to the input requirements of the YOLO-v7 algorithm, the image is cropped and scaled, and the target light language area is adjusted to an appropriate size to ensure that the light language signal can be correctly and effectively recognized by the YOLO-v7 model; the image is converted from the RGB color space to the HSV or YUV color space. Through this conversion, the color characteristics of the light language signal under different lighting conditions are enhanced, so that it can be accurately recognized in different environments.

[0025] S3. Use the YOLO-v7 algorithm to analyze the pre-processed light image to identify the target HDV's various driving intentions, including the vehicle's turning intention, braking intention, and overtaking intention. This process involves multiple steps of image feature extraction, candidate region generation and optimization, target classification and confidence prediction, and finally infers the HDV's specific driving behavior through comprehensive analysis. The steps are:

[0026] Step 1: Obtain all the light language images of historical manually driven vehicles to obtain a training set;

[0027] Step 2: Preprocess the light language images in the training set, convert the RGB images into YUV or HSV color space, and use image enhancement technology to improve the image quality;

[0028] Step 3: Input the preprocessed light signal image into the optimized YOLO-v7 network, and extract multi-scale features in the image through the convolutional neural network to improve the accuracy and adaptability of recognition. Specifically, the model can extract spatial and contextual features of light signal from different scales by enhancing the structure of the convolutional layer; increase the depth of the convolutional layer and improve the efficiency of feature extraction to ensure that the changes of light signal can be accurately captured in complex traffic environments;

[0029] Step 4: Use the region proposal network RPN to process multi-scale features, introduce adaptive candidate box generation technology, dynamically generate more appropriate candidate box sizes and generate candidate regions according to actual traffic conditions and historical data;

[0030] Step 5: classify each candidate area and identify whether there is a light signal in the candidate area. If yes, execute step 6; if no, discard the candidate area.

[0031] Step 6: Identify the driving status of the manually driven vehicle according to the light signal, classify the light signal, and identify the driving intention;

[0032] Step 7: A weighted strategy is introduced into the loss function of the YOLO-v7 model, that is, the classification loss, bounding box positioning loss, and confidence loss are weighted and optimized, and their weights are adjusted dynamically. The loss function of the model is calculated, and the L2 regularization term is introduced to prevent the model from overfitting during training. The Adam optimization algorithm is used to optimize the model parameters, and the model training is completed when the loss function converges.

[0033] Specifically, the preprocessed light language image is input into the YOLO-v7 model. The input preprocessed image is firstly subjected to feature extraction by the YOLO-v7 deep convolutional neural network. This step extracts the spatial features and semantic features of the image step by step through multiple convolutional layers:

[0034] G i =φ(W i ·X+c i )

[0035] Among them, G i represents the output feature map of the i-th layer; W i and c i are the weight and bias of the i-th layer respectively; X is the input image or the output of the previous layer; φ is the activation function (ReLU).

[0036] Each convolution layer forms a multi-scale feature map by continuously extracting low-level to high-level features. After a series of convolution operations, the feature map becomes more refined and abstract, and these feature maps will be used for subsequent region proposals and object detection:

[0037] G i+1 =φ(W i+1 ·G i +c i+1 )

[0038] As the network depth increases, the spatial resolution of the feature map decreases, but the semantic information increases. After feature extraction, the obtained feature map contains rich spatial and semantic information in the image.

[0039] After the feature map is extracted, the RPN network of YOLO-v7 generates multiple anchor boxes based on the feature map of the previous layer. These anchor boxes represent the locations where the target objects may exist in the image. Each anchor box will be classified and regressed to determine whether it contains the target:

[0040]

[0041]

[0042]

[0043]

[0044] Among them, Δx is the horizontal coordinate offset of the prediction box; x is the horizontal coordinate of the center of the prediction box of the target object; x anchor is the horizontal coordinate of the center of the anchor box; w anchor The width of the anchor box; Δy is the vertical coordinate offset of the prediction box; y is the vertical coordinate of the center of the prediction box of the target object; y anchor is the vertical coordinate of the center of the anchor box; h anchor is the height of the anchor box; Δw is the normalized logarithmic offset of the predicted box width relative to the anchor box; w pred is the predicted box width of the target object; Δh is the normalized logarithmic offset of the predicted box height relative to the anchor box; h pred is the predicted box height of the target object; h anchor is the height of the anchor box.

[0045] The generated anchor frame will be further optimized through regression analysis to fit the location of the target light signal more accurately. This optimization is achieved through the regression loss function:

[0046]

[0047] Among them, L regis the regression loss function, which measures the position and size difference between the predicted box and the real box; N pos is the number of positive samples, that is, the number of samples correctly identified as targets; Δx i , Δy i , Δw i , Δh i is the regression offset of the prediction box, which represents the center coordinate offset and the adjustment of width and height of the prediction box respectively; is the corresponding regression parameter of the actual target box, which is used to compare with the predicted value. The model will be used multiple times in each iteration during training until the final loss function value converges or is small enough, indicating that the position of the predicted box is accurate enough.

[0048] Based on the optimized anchor frame, YOLO-v7 further accurately locates the position of the light signal through bounding box regression. The regression formula is as follows:

[0049] BBox(x,y,w,h)=(φ(Δx)+x anchor ,φ(Δy)+y anchor ,e Δw ·w anchor ,e Δh ·h anchor )

[0050] Among them, BBox(x,y,w,h) is the final predicted bounding box, expressed as the center coordinates (x,y) of the bounding box and the width w and height h; φ(Δx)+x anchor and φ(Δy)+y anchor are the final horizontal and vertical coordinates of the center point of the prediction box; φ is the activation function (ReLU); x anchor and anchor are the center coordinates of the anchor box; e Δw ·w anchor and e Δh ·h anchor is the width and height of the prediction box, multiplied by the width w of the anchor box after exponential scaling anchor and height h anchor Thus, the final size is obtained. The bounding box not only indicates the location of the light signal, but also includes its size information, and the position and size of the anchor box are adjusted to be closer to the position and size of the actual target.

[0051] The mean square error (MSE)-based regression method is used to optimize the center coordinates and size errors of the bounding box. The bounding box positioning loss expression is:

[0052]

[0053] Among them, Lcoord The mean square error (MSE) of bounding box positioning is used to evaluate the difference between the predicted bounding box and the actual bounding box; S 2 is the number of grids into which the image is divided; B is the number of bounding boxes predicted for each grid; is the indicator function, which takes the value of 1 when the jth anchor box in the i-th unit contains the target object, otherwise it takes the value of 0; i ,y i is the center coordinate of the actual bounding box; is the center coordinate of the predicted bounding box; w i ,h i is the width and height of the actual bounding box; is the width and height of the predicted bounding box.

[0054] For each generated bounding box, YOLO-v7 classifies the target through the Softmax function to determine which light signal it belongs to (such as turning, braking, overtaking, etc.). The classification formula is:

[0055]

[0056] Among them, η i is the prediction score of category i; k is the total number of categories; σ(η i ) is the probability of the i-th category, calculated by the Softmax function.

[0057] Classification loss L cls It is used to measure the classification accuracy of the model in the driving intention recognition task. Specifically, the loss function evaluates the classification ability of the model by calculating the weighted negative log-likelihood value between the predicted label and the true label. That is, the category weight α is introduced c , so that different categories of light language signals can be given different weights according to their frequency or importance, solving the problem of data imbalance, which is expressed as follows:

[0058]

[0059] Among them, α c is the class weight, which is used to deal with the class imbalance problem; y ic Indicates that the i-th sample belongs to the true label of the c-th category (the value is 0 or 1); is the probability that the ith sample is predicted to be the cth class after adjustment by the activation function; N is the total number of samples; c is the total number of categories. The classification loss is calculated by introducing category weights and activation functions to handle category imbalance during model training and optimize the model's classification accuracy for various driving intentions (such as steering, braking, overtaking, etc.).

[0060] In addition to classification, YOLO-v7 also needs to predict the confidence that each bounding box contains the target object. The calculation of confidence determines the model's degree of certainty about the existence of the light signal. Introducing the penalty coefficient λ in the absence of an object noobj , reduce the network's misdetection of invalid targets, and the confidence loss is calculated as:

[0061]

[0062] Among them, L conf is the confidence loss function, which is used to evaluate the difference between the confidence that the target object is contained in the predicted bounding box and the actual situation; C i is the predicted confidence score, indicating the probability that the target object exists in the bounding box; is the actual confidence level (ideally 1 or 0); is an indicator function, which takes the value of 1 when the anchor box contains the target object; is an indicator function, which takes the value of 1 when the anchor box does not contain the target object; noobj It is the penalty coefficient in the absence of an object, which is used to reduce the impact of confidence errors on irrelevant areas.

[0063] After completing the classification and confidence prediction, YOLO-v7 combines the location, category, and confidence information of the light signal to infer the HDV's driving intention. By combining the classification results of multiple candidate boxes, the system infers the HDV's current driving behavior, such as braking, steering, or overtaking. The formula for intention analysis is as follows:

[0064]

[0065] Where I represents the comprehensive driving intention score; N is the total number of candidate boxes; p i is the predicted probability of the i-th candidate box; They are indicator functions, indicating whether the candidate box represents a light signal for turning, braking or overtaking; L turn , L brake , L overtake are the feature confidences of steering, braking and overtaking signals respectively.

[0066] The YOLO-v7 algorithm combines classification loss, confidence loss, and bounding box localization loss into an overall loss function and optimizes it through back propagation. This process ensures the accuracy and reliability of the model in different driving intention recognition tasks.

[0067] In order to enhance the generalization ability of the model, a regularization term based on the L2 norm is added to the loss function to control the complexity of the model. The final optimized loss function is:

[0068] L=λcoord L coord +λ conf L conf +λ cls L cls +λ reg ·W 2

[0069] Among them, λ coord is the weight coefficient of the bounding box positioning loss; L coord is the bounding box positioning loss, which measures the error of the bounding box position and size; conf is the weight coefficient of confidence loss; L conf is the confidence loss, which measures the probability error of the target; cls is the weight coefficient of classification loss; L cls is the classification loss, which measures the accuracy of target classification; reg is the weight of the regularization term; W 2 The L2 norm of the model parameters is used to reduce the risk of overfitting. Finally, the Adam optimization algorithm is used to continuously adjust the model parameters to minimize the loss function and improve the recognition performance of the model.

[0070] S4. Through the above steps, the driving intentions (steering, braking, overtaking) of the HDV have been identified by the YOLO-v7 algorithm. In order to ensure that the CAV can respond in time, these identified driving intentions are transmitted to the OBU terminal of the CAV in real time through the 5G communication network. Taking advantage of the high bandwidth and low latency characteristics of the 5G network, the identified driving intentions are transmitted to the target CAV in a unicast manner to ensure the timeliness and reliability of data transmission.

[0071] In this embodiment, after the OBU terminal of CAV receives the driving intention information of HDV, it first analyzes the data. CAV combines the received driving intention and its own sensor data, and uses the built-in decision support system to analyze the current environment. The decision support system considers multiple factors: the intention of HDV, the traffic conditions of the current lane, the traffic volume of the adjacent lanes, and the driving status of CAV itself. Based on these analysis results, CAV generates multiple feasible driving strategies and selects the optimal strategy through the path planning algorithm. After selecting the optimal driving strategy, the control system of CAV generates specific driving behavior instructions according to the strategy. These instructions include deceleration, lane change, acceleration or emergency braking, etc. The control system makes corresponding adjustments to the vehicle's power system, steering system and braking system according to these instructions to ensure that CAV drives safely according to the optimal decision. During the execution of driving behavior, the sensor system of CAV continuously monitors the environment around the vehicle and its own status. If the environment changes, such as the sudden deceleration or lane change of the vehicle in front, CAV will immediately re-evaluate the current situation and adjust the driving strategy through the decision support system to ensure the safe driving of the vehicle. At the same time, the execution results will be transmitted to the CAV's control system and vehicle networking system through a feedback mechanism to optimize future decision-making processes.

[0072] In this embodiment, a corresponding driving strategy and a feedback mechanism are formulated, and the intelligent connected vehicle (OBU) sends feedback to the MEC node based on the received driving intention information and actual driving conditions for continuous learning and optimization of the model.

[0073] A HDV driving intention recognition system based on light language in mixed traffic flow, the system comprises: a real-time light language image acquisition unit, an HDV driving intention calculation module, a vehicle network intention sharing service module and a driving behavior decision execution unit. These units work together to achieve real-time recognition and response to HDV driving intention in mixed traffic flow.

[0074] The real-time light language image acquisition unit is composed of an on-board visual sensor installed on a CAV (connected autonomous vehicle). The unit uses a high-definition camera to collect light language signal images of surrounding HDVs (manually driven vehicles) in real time. These visual sensors can accurately capture the light language information of HDVs under different lighting conditions (such as daytime, nighttime, rainy and foggy weather), and transmit these image data in real time to the on-board computing unit for preliminary processing. The collected light language image data is transmitted to the MEC (edge ​​computing) node via the 5G communication network to ensure that the data can be transmitted quickly and stably in a complex traffic environment.

[0075] The HDV driving intention calculation module is deployed in the MEC node. The module first preprocesses the received light language image data, including image enhancement, denoising, cropping, and color space conversion to improve the image quality and the recognition accuracy of the light language signal. Then, the YOLO-v7 algorithm is used to extract and analyze the features of the preprocessed image to identify the specific driving intentions of the HDV, such as the vehicle's steering, braking, and overtaking intentions. Through the processing of this module, the system can efficiently and accurately identify the driving behavior of the HDV, providing reliable data support for the decision-making of the CAV.

[0076] The Internet of Vehicles intention sharing service module unicasts the identified HDV driving intention information to the OBU (on-board unit) terminal of the target CAV through the 5G communication network. This module uses the high bandwidth and low latency characteristics of the 5G network to ensure that the driving intention information can be delivered to the target CAV in the shortest time. Through this module, the CAV can receive the driving intentions of the surrounding HDVs in real time, and make corresponding driving decisions based on the received information, ensuring safety and efficiency in complex traffic environments.

[0077] The driving behavior decision execution unit is located inside the CAV. The unit receives the HDV driving intention information from the OBU, and combines the vehicle's own sensor data and current traffic conditions to perform a comprehensive analysis using the decision support system. Based on the analysis results, the unit generates the optimal driving behavior strategy, such as deceleration, acceleration, lane change or emergency braking. These driving instructions are executed by the CAV's control system to ensure that the vehicle drives safely according to the optimal strategy. In addition, the unit also has a dynamic adjustment function, which can adjust the driving behavior according to real-time environmental changes to cope with emergencies, further improving the CAV's ability to cope with mixed traffic flows.

[0078] In a specific embodiment of the present invention, Figure 3As shown in the figure, CAV (intelligent connected vehicle) collects the light signal images of surrounding HDV (manually driven vehicles) in real time through on-board visual sensors. When the light signal of the target HDV is successfully detected, the system transmits the collected image data to the MEC (edge ​​computing) node through the 5G communication network. At the MEC node, the image data is first preprocessed, including denoising, image enhancement and other steps to improve the quality of the light signal and the accuracy of subsequent recognition. The preprocessed image data is then used for driving intention recognition through the YOLO-v7 algorithm. The algorithm can identify the specific driving intention of the HDV, including steering intention, braking intention and overtaking intention. The identified driving intention is transmitted to the OBU of the target CAV through 5G network unicast. After receiving the intention, the OBU makes corresponding driving decisions, including changing lanes, slowing down or overtaking, in combination with the current environmental information. Finally, the CAV performs the corresponding driving operations and shares driving information with surrounding vehicles through the collaborative driving system, thereby improving the safety and efficiency of traffic. The entire system is efficiently processed at the MEC node, which can quickly respond to and adjust the driving behavior of the CAV, and realize intelligent perception and decision-making of complex traffic environments. The overall schematic diagram is as follows Figure 4 shown.

[0079] The system in the embodiment of the present invention can effectively identify and share the driving intention of HDVs in mixed traffic flow, and improve the overall safety and efficiency of traffic through intelligent decision-making and execution of CAVs.

[0080] Table 1 YOLO-v7 parameter settings

[0081]

[0082]

[0083] Table 2 Training and testing dataset settings

[0084]

[0085] Based on the model parameters and the training and test data set settings, the recognition results shown in Table 3 can be obtained:

[0086] Table 3 Model evaluation indicators

[0087]

[0088] In Table 3, the average accuracy of the model at the 0.5IoU threshold is 68.5%, and the average accuracy is 45.7% within the stricter multiple IoU threshold ranges. These results show that the model has high accuracy in locating targets, especially at lower IoU thresholds. The model has an accuracy of 91.3% and a recall rate of 89.5%, which shows that in actual traffic scenarios, the model can accurately and effectively identify most driving intentions. In addition, the inference time of 15ms / frame ensures that the system can respond in real time, which is of great significance for the practical application of intelligent connected vehicles. This patent embodiment successfully implements an efficient driving intention recognition system by optimizing and adjusting the YOLO-v7 model. The model shows strong recognition capabilities in complex traffic flow environments, and can provide timely and accurate driving intention information for intelligent connected vehicles, supporting them to make safe and reasonable driving decisions. These evaluation results not only meet the actual application needs, but also prove the practical application value of the present invention in intelligent transportation systems.

[0089] The system implementation of the present invention is the same as the method implementation.

[0090] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0091] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying HDV driving intention based on light language in mixed traffic flow, characterized in that: include: An autonomous driving network is constructed, wherein the autonomous driving network includes a connected autonomous driving vehicle CAV, a manually driven vehicle HDV, and a MEC node; the CAV obtains the light language images of the surrounding target HDV; the light language images are transmitted to the MEC node, and the MEC node preprocesses the light language images; the trained driving intention recognition model based on the YOLO-v7 network is used to recognize the driving intention of the target HDV on the preprocessed light language images; the HDV driving intention is transmitted to the CAV; the CAV generates corresponding driving decisions according to the HDV driving intention and executes the driving decisions.

2. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 1 is characterized in that: The light language images of the target HDV include brake lights, turn signals, and overtaking signal lights.

3. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 1, characterized in that: The light language image is transmitted to the MEC node in real time through 5G communication.

4. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 1, characterized in that: The preprocessing of the light language image includes: enhancing the light language image by using histogram equalization and gamma correction technology; performing median filtering on the enhanced light language image; cropping and scaling the filtered image; converting the RGB color space of the cropped and scaled image into the HSV color space or the YUV color space to obtain the preprocessed light language image.

5. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 1 is characterized in that: Training the driving intention recognition model based on the YOLO-v7 network includes: Step 1: Obtain all the light language images of historical manually driven vehicles to obtain a training set; Step 2: Preprocess the light language images in the training set, that is, convert the RGB images into YUV or HSV color space, and use image enhancement technology to enhance the images in the converted color space; Step 3: Input the preprocessed light language image into the optimized YOLO-v7 network, and extract multi-scale features in the image through the convolutional neural network; Step 4: Use the region proposal network (RPN) to process multi-scale features, that is, introduce adaptive candidate box generation technology to dynamically generate candidate boxes of corresponding sizes based on actual traffic conditions and historical data; and select candidate regions based on the candidate boxes; Step 5: classify each candidate area and identify whether there is a light signal in the candidate area. If yes, execute step 6; if no, discard the candidate area. Step 6: Identify the driving status of the manually driven vehicle according to the light signal, classify the light signal, and identify the driving intention; Step 7. Calculate the loss function of the YOLO-v7 model based on the driving intention. A weighted strategy is introduced into the loss function, that is, the classification loss, bounding box positioning loss, and confidence loss are weighted and optimized, and their weights are adjusted dynamically. An L2 regularization term is introduced into the loss function. The Adam optimization algorithm is used to optimize the model parameters. When the loss function converges, the model training is completed.

6. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 5 is characterized in that: The convolutional neural network extracts multi-scale features in the image as follows: G i =φ(W i ·X+c i ) G i+1 =φ(W i+1 ·G i +c i+1 ) Among them, G i represents the output feature map of the i-th layer; W i and c i are the weight and bias of the i-th layer respectively; X is the input image or the output of the previous layer; φ is the activation function.

7. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 5 is characterized in that: The processing of multi-scale features using the region proposal network RPN includes: generating multiple anchor boxes of different sizes and aspect ratios for each pixel in the feature map; for each anchor box, predicting the probability and category of the target it contains, and predicting the precise bounding box coordinates of the target; the formula is: Among them, Δx is the horizontal coordinate offset of the prediction box; x is the horizontal coordinate of the center of the prediction box of the target object; x anchor is the horizontal coordinate of the center of the anchor box; w anchor is the width of the anchor box; Δy is the vertical coordinate offset of the prediction box; y is the vertical coordinate of the center of the prediction box of the target object; y anchor is the vertical coordinate of the center of the anchor box; h anchor is the height of the anchor box; Δw is the normalized logarithmic offset of the predicted box width relative to the anchor box; w pred is the predicted box width of the target object; Δh is the normalized logarithmic offset of the predicted box height relative to the anchor box; h pred is the predicted box height of the target object; h anchor is the height of the anchor box.

8. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 5 is characterized in that: Identifying the driving state of the manually driven vehicle includes calculating the driving intention confidence of the manually driven vehicle, and obtaining the driving intention of the manually driven vehicle according to the confidence value; Its calculation expression is: Where I represents the comprehensive driving intention score; N is the total number of candidate boxes; p i is the predicted probability of the i-th candidate box; They are indicator functions, indicating whether the candidate box represents a light signal for turning, braking or overtaking; L turn , L brake , L overtake are the feature confidences of steering, braking and overtaking signals respectively.

9. The method for identifying HDV driving intention based on light language in mixed traffic flow according to claim 5, characterized in that: The loss function of the model includes bounding box localization loss, confidence loss and classification loss; The mean square error-based regression method is used to optimize the center coordinates and size errors of the bounding box. That is, the bounding box positioning loss expression is: Among them, L coord is the mean square error of bounding box positioning; S 2 is the number of grids into which the image is divided; B is the number of bounding boxes predicted for each grid; is the indicator function; x i ,y i is the center coordinate of the actual bounding box; is the center coordinate of the predicted bounding box; w i ,h i is the width and height of the actual bounding box; To predict the width and height of the bounding box; Introducing the penalty coefficient λ for the case where there is no object noobj , reducing the network's misdetection of invalid targets, where the confidence loss expression is: Among them, L conf is the confidence loss function; C i is the confidence score of the prediction; is the actual confidence level; are indicator functions; noobj is the penalty coefficient when there is no object; Introducing category weight α c , different weights are assigned according to the frequency or importance of different categories of light language signals, where the classification loss expression is: Among them, α c is the class weight; y ic Indicates that the i-th sample belongs to the true label of the c-th category; is the probability that the i-th sample is predicted to be the c-th class after adjustment by the activation function; N represents the total number of samples; c represents the total number of categories; The optimized model loss function is: L=λ coord L coord +λ conf L conf +λ cls L cls +λ reg ·||W|| 2 Among them, λ coord is the weight coefficient of the bounding box positioning loss; L coord is the bounding box localization loss; λ conf is the weight coefficient of confidence loss; L conf is the confidence loss; cls is the weight coefficient of classification loss; L cls is the classification loss; reg is the weight of the regularization term; ||W|| 2 Represents the L2 norm of the model parameters.

10. A system for identifying HDV driving intention based on light language in mixed traffic flow, the system being used to execute a method for identifying HDV driving intention based on light language in mixed traffic flow as claimed in any one of claims 1 to 9, characterized in that: include: Real-time light language image acquisition unit, HDV driving intention recognition module, Internet of Vehicles intention sharing service module and driving behavior decision execution unit; The real-time light language image acquisition unit is used to acquire HDV light language images; The HDV driving intention recognition module recognizes the driving intention of the HDV based on the collected light language image and the driving intention recognition model based on the YOLO-v7 network on the MEC node; The Internet of Vehicles intention sharing service module unicasts the driving intention of the HDV to the target CAV; The driving behavior decision execution unit makes decisions and plans paths according to the received HDV driving intention information, and executes corresponding driving behaviors.

Images