A real-time evaluation method and system for driving behavior based on Yolo and TransTrack

Abstract

The invention discloses a real-time evaluation method and system for driving behavior based on Yolo and TransTrack, comprising: obtaining a road scene video stream in real time; using a YOLO model to perform target detection on each frame image in the road scene video stream to obtain a target detection result one; using the target detection result one as a priori feature input of a TransTrack model, the TransTrack model combines each frame image in the road scene video stream with target feature information corresponding to the frame image, performs a target tracking task, and forms target tracking data; using a visual recognition model to extract current traffic scene features from each frame image in the road scene video stream; determining a current driving behavior type based on the current traffic scene features and the current target tracking data; judging whether the current driving behavior type complies with traffic regulations based on the current driving behavior type to obtain a behavior analysis result; and scoring and warning the current driving behavior based on the behavior analysis result.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent connected vehicle road testing, and specifically relates to a real-time driving behavior evaluation method and system based on Yolo and TransTrack. Background Art

[0002] With the rapid development of intelligent transportation systems and autonomous driving technology, accurate evaluation of driving behavior has become a key technology to improve road safety and driving experience. Traditional driving behavior analysis methods rely on manual observation and post-data processing, which is not only time-consuming and inefficient, but also has limitations in real-time and accuracy. In order to solve these problems, researchers and engineers have been exploring more efficient and automated solutions.

[0003] Among the existing methods, rule-based driving behavior evaluation methods are a common practice. This driving behavior evaluation method uses a predefined set of rules to determine whether the driving behavior meets safety standards. For example, speeding behavior is identified by setting a speed threshold, or a lane line detection algorithm is used to evaluate whether the vehicle deviates from the lane. Although this method is relatively simple to implement and easy to understand, it has limited flexibility and is difficult to adapt to changing traffic environments and complex driving scenarios. In addition, with the emergence of new types of driving behaviors, the rule base needs to be continuously updated and maintained, which increases the complexity and maintenance cost of the system.

[0004] The second category is the driving behavior evaluation method based on machine learning. This driving behavior evaluation method analyzes data collected from sensors and cameras through training algorithms, such as support vector machines (SVM) or random forests, to identify and evaluate driving behavior. Machine learning methods are able to process large amounts of data and learn complex patterns from them. As the amount of data increases, the accuracy of the evaluation can also be gradually improved. However, this method requires a large amount of labeled data for training, and has high requirements for data quality. At the same time, the model has poor interpretability, which makes it difficult to understand the specific decision-making process of the model, which may sometimes be detrimental to safety-critical applications.

[0005] The third category is the driving behavior evaluation method based on deep learning, which uses deep neural networks, such as convolutional neural networks (CNN) or recurrent neural networks (RNN), to perform deep feature extraction and behavior analysis on video and sensor data. These methods perform well in processing image and sequence data, especially for complex driving scenarios. Although deep learning methods have advantages in accuracy and adaptability, they usually require a lot of computing resources and training time. In addition, the model may be too complex, which may lead to overfitting and affect its generalization ability in practical applications. Summary of the invention

[0006] Purpose of the invention: To solve one of the technical problems existing in the prior art, the present invention proposes a real-time evaluation method and system for driving behavior based on Yolo and TransTrack, which utilizes target detection and tracking technology to achieve real-time evaluation and analysis of driving behavior.

[0007] Technical solution: A real-time evaluation method for driving behavior based on YOLO and TransTrack, including the following steps:

[0008] Obtain road scene video stream in real time;

[0009] The YOLO model is used to perform target detection on each frame of the road scene video stream to obtain target detection result 1;

[0010] The target detection result is used as a priori feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data; the target tracking data includes: target position change data, target size change data, and target direction change data;

[0011] Using the visual recognition model, the current traffic scene features are extracted from each frame of the road scene video stream;

[0012] According to the current traffic scene characteristics and the current target tracking data, the current driving behavior type is determined;

[0013] Based on the current driving behavior type, determine whether it complies with traffic regulations and obtain behavior analysis results;

[0014] Based on the behavior analysis results, the current driving behavior is scored and warned.

[0015] Furthermore, the target detection result is used as a priori feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data. The specific operations include:

[0016] The TransTrack model performs target detection on each frame of the road scene video stream and obtains target detection result 2;

[0017] A weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model;

[0018] Using the adjusted confidence, the target detection result 1 and the target detection result 2 are fused to determine the final target detection result;

[0019] The spatial relationship is used to associate the final target detection results and the tracking trajectory of the TransTrack model to form target tracking data.

[0020] Furthermore, the weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model. The specific operations include:

[0021] Determine the evaluation metrics and evaluation metric functions used to measure the performance of the YOLO module and TransTrack model;

[0022] According to the evaluation indicators, the performance data of the YOLO model and TransTrack model are collected in real time;

[0023] According to the following formula, the confidence weight of the YOLO model and the confidence weight of the TransTrack model are dynamically adjusted according to the performance data of the YOLO model and the TransTrack model:

[0024] w YOLO =f(YOLO Performance Metrics)

[0025] w TransTrack =g(TransTrack Performance Metrics)

[0026] Where f(YOLO Performance Metrics) and g(TransTrack Performance Metrics) represent the evaluation index functions of the YOLO model and the TransTrack model respectively; w YOLO Represents the confidence weight of the YOLO model, w TransTrack Represents the confidence weight of the TransTrack model;

[0027] According to the adjusted confidence weights, the confidence of the YOLO model and the confidence of the TransTrack model are determined.

[0028] Furthermore, the evaluation indicators include accuracy, recall, average precision and tracking accuracy.

[0029] Furthermore, the warning includes: sound warning, visual warning and tactile warning; the tactile warning includes seat vibration or seat belt tightening.

[0030] Furthermore, the method further comprises the following steps:

[0031] Collect current driving decision information, adjust the parameters of the YOLO model and the TransTrack model based on the current driving decision information, the current traffic scene information and the target tracking data, and obtain the updated YOLO model and the TransTrack model;

[0032] The updated YOLO model and TransTrack model are used to perform target detection and target tracking on each frame image in the road scene video stream.

[0033] The present invention discloses a driving behavior real-time evaluation system based on YOLO and TransTrack, comprising:

[0034] A video stream acquisition module is used to acquire road scene video streams in real time;

[0035] The target detection module is used to use the YOLO model to perform target detection on each frame of the road scene video stream to obtain a target detection result 1;

[0036] The target tracking module is used to use the target detection result as the prior feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data; the target tracking data includes: target position change data, target size change data, and target direction change data;

[0037] The driving behavior type recognition module is used to extract the current traffic scene features from each frame image in the road scene video stream using a visual recognition model; and determine the current driving behavior type based on the current traffic scene features and the current target tracking data;

[0038] The driving behavior analysis module is used to determine whether the current driving behavior type complies with traffic regulations and obtain the behavior analysis results;

[0039] The warning module is used to score and warn the current driving behavior based on the behavior analysis results.

[0040] Furthermore, the target detection result is used as a priori feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data. The specific operations include:

[0041] The TransTrack model performs target detection on each frame of the road scene video stream and obtains target detection result 2;

[0042] A weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model;

[0043] Using the adjusted confidence, the target detection result 1 and the target detection result 2 are fused to determine the final target detection result;

[0044] The spatial relationship is used to associate the final target detection results and the tracking trajectory of the TransTrack model to form target tracking data.

[0045] Furthermore, the weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model. The specific operations include:

[0046] Determine the evaluation metrics and evaluation metric functions used to measure the performance of the YOLO module and TransTrack model;

[0047] According to the evaluation indicators, the performance data of the YOLO model and TransTrack model are collected in real time;

[0048] According to the following formula, the confidence weight of the YOLO model and the confidence weight of the TransTrack model are dynamically adjusted according to the performance data of the YOLO model and the TransTrack model:

[0049] w YOLO =f(YOLO Performance Metrics)

[0050] w TransTrack =g(TransTrack Performance Metrics)

[0051] Where f(YOLO Performance Metrics) and g(TransTrack Performance Metrics) represent the evaluation index functions of the YOLO model and the TransTrack model respectively; w YOLO Represents the confidence weight of the YOLO model, w TransTrack Represents the confidence weight of the TransTrack model;

[0052] According to the adjusted confidence weights, the confidence of the YOLO model and the confidence of the TransTrack model are determined.

[0053] Furthermore, it also includes a model optimization module;

[0054] The model optimization module is used to collect current driving decision information, adjust the parameters of the YOLO model and the TransTrack model based on the current driving decision information, the current traffic scene information and the target tracking data, and obtain the updated YOLO model and the TransTrack model.

[0055] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0056] (1) The method of the present invention can effectively identify and judge the behavior pattern of the execution of the intelligent networked vehicle algorithm, and provide instant feedback and improvement suggestions to the safety officer by real-time monitoring and evaluating the driver's behavior. On the one hand, the method of the present invention can determine whether the intelligent driving behavior complies with traffic regulations, and on the other hand, improve traffic safety during the test process; it realizes efficient and accurate real-time evaluation of the driving behavior driven by the intelligent driving algorithm, and improves road traffic safety during the road test;

[0057] (2) The method of the present invention can make full use of the information of target detection and target tracking to achieve efficient and accurate estimation of driving behavior;

[0058] (3) The method of the present invention performs adaptive learning based on the driver's behavior pattern and can continuously learn from new driving behavior data to adapt to the ever-changing traffic environment and driving behavior, continuously optimize, and provide long-term behavior improvement suggestions. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of a real-time evaluation method for driving behavior based on Yolo and TransTrack disclosed in the present invention;

[0060] Figure 2 A schematic diagram of the driving behavior evaluation process using an intersection with traffic lights as a scenario. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is now further described in conjunction with the accompanying drawings and embodiments.

[0062] Embodiment 1:

[0063] like Figure 1 As shown, this embodiment discloses a real-time evaluation method for driving behavior based on YOLO target detection and TransTrack target tracking technology, which mainly includes the following steps:

[0064] Step 1: Capture a real-time road scene video stream through a vehicle camera or other video source, and pre-process the real-time road scene video stream to improve image processing accuracy;

[0065] Step 2: Use the YOLO model to detect the target in each frame of the video stream to obtain target feature information, which includes but is not limited to the bounding box, confidence, target position, target movement direction and target category; the specific operations of this step include:

[0066] The pre-trained YOLO model divides each frame of the video stream into S×S grid cells; each grid cell is responsible for detecting whether the center of the target is located within the grid cell and calculating the deep learning features of the target; based on the deep learning features of the target, the target is classified and the target feature information is output;

[0067] Step 3: Use the target feature information obtained in step 2 as the prior feature input of the TransTrack model. After obtaining the prior feature input, the TransTrack model combines this prior information to perform the target tracking task, handle occlusion, target generation and disappearance, and obtain tracking data. The tracking data includes: the position change, size change and direction change of the target. The specific operations include:

[0068] For each detected target, the TransTrack model generates a tracking identifier; based on the target's motion and appearance changes, the TransTrack model updates the tracking identifier and maintains the target's identity information; when the target is occluded or temporarily leaves the field of view, the TransTrack model maintains tracking of the target until the target reappears in the field of view.

[0069] When using the TransTrack model alone, if the initial features are inaccurate, the tracking algorithm may not be able to correctly identify and track the target. In addition, although the TransTrack model can continue tracking after occlusion, the premise is that the trajectory trend or speed of the occluded vehicle has not changed significantly. This situation is similar to the recapture process after the target is lost, and this process is difficult to track and match the tracking trajectory of the original target in the absence of semantic support.

[0070] To this end, before the TransTrack data is imported, this embodiment uses the YOLO model to perform pre-fusion of target feature data, and the target feature information (such as bounding box, confidence, object attributes, position, direction, speed, etc.) detected by the YOLO model is used as the initial feature input of the TransTrack model, that is, the input of the initialized TransTrack is not only the video stream image frame, but also includes the initial position (pixel coordinates), target size, direction, and other feature information input by the YOLO model outside the image frame, while realizing seamless data flow, while enhancing the target representation ability. Especially in the dynamically changing traffic flow scene, in response to vehicle acceleration, deceleration, lane change and other behaviors, as well as occlusion, reproduction and other processes, and even tracking loss caused by changes in illumination angles, the semantic understanding function of the YOLO model is used to dynamically update the bounding box, confidence, object attributes, position, direction and other information, thereby realizing automatic adjustment of the input features of the TransTrack model. In this process, the target is not simply identified using the YOLO model and then imported into the TransTrack model for tracking. Instead, the feature information of the target identified by the YOLO model is extracted and used as the prior feature input of the TransTrack model. After obtaining the prior feature input, the TransTrack model combines this prior information to perform the target tracking task. Therefore, step 2 is mainly used to solve the defect of inaccurate initial features when the TransTrack model is used independently as described above.

[0071] If the YOLO model and the TransTrack model are simply superimposed, since the TransTrack model itself also contains the function of target recognition, there will be inconsistent judgments on the same target during the decision fusion stage between the two. This embodiment uses a weighted adaptive adjustment method to adjust the confidence of the detection and tracking algorithm. Then, the spatial relationship (distance, direction, and movement direction trend) is used to associate the detection results of the YOLO model and the tracking trajectory of the TransTrack model. In the adaptive adjustment process, an adaptive factor is introduced to evaluate the prior calculation results.

[0072] The overall idea of ​​adaptive adjustment in this embodiment is to estimate and correct the model parameters and system noise in real time during the calculation process to improve the algorithm accuracy and reduce the state estimation error. The adaptive factor is introduced in the iteration using an iterative method. There is an optimal adaptive factor in each iteration, which is combined with the statistical analysis method for residuals such as variance component estimation to achieve the final convergence effect, thereby confirming relatively reliable confidence parameters. Parameters such as variance component estimation and residuals are statistical methods commonly used in data processing. The use process will be accompanied by statistics of the connection algorithm, and the statistical parameters of the current cycle will be generated according to the process of each cycle, so as to realize the judgment of the convergence effect. So it is actually the statistics realized by the algorithm accompanying the adjustment process.

[0073] The principle of adaptation is to add an adjustment parameter, which is called a factor. The calculation of the factor is realized through a piecewise function. The introduction of the adaptive factor will adjust the noise estimation in real time, and then compensate and correct it according to the estimated noise.

[0074] With the confidence level of the connection between the two models as a reference, fine-tuning parameters are added when confirming the credibility between the two models. These fine-tuning parameters were adjusted using empirical values ​​in the past in other similar model parameter adjustment processes, but this embodiment introduces the idea of ​​adaptive adjustment for the accuracy adjustment when Yolo and TransTrack are used together. The model parameters that can be adaptively adjusted using the adaptive factor method include: tracking threshold, search window size, feature weight, etc.

[0075] The following is a further explanation of the weighted adaptive adjustment method for adjusting the confidence of Yolo and TransTrack in target recognition. The following steps are included:

[0076] 1. Define evaluation metrics: Define evaluation metrics to measure the performance of YOLO and TransTrack. These metrics include but are not limited to:

[0077] Precision: The proportion of correctly detected / tracked objects.

[0078] Recall: The ratio of detected / tracked targets to all actual targets.

[0079] Average Precision (AP): The average precision at different confidence thresholds.

[0080] Tracking Accuracy: The consistency between the tracked target and the true target trajectory.

[0081] 2. Collect performance data: Run YOLO and TransTrack in real scenarios and collect their performance data in real time based on evaluation indicators. The performance data includes but is not limited to: detection and tracking accuracy, stability, and response time.

[0082] 3. Design weighting function: Dynamically adjust the confidence weights of YOLO and TransTrack according to their performance data, expressed as:

[0083] w YOLO =f(YOLO Performance Metrics)

[0084] w TransTrack =g(TransTrack Performance Metrics)

[0085] Where f(YOLO Performance Metrics) and g(TransTrack Performance Metrics) represent functions defined according to performance indicators, which map performance indicators to weight values; w YOLO Represents the confidence weight of YOLO, w TransTrack Represents the confidence weight of TransTrack;

[0086] 4. Determine the weight adjustment strategy so that when the detection and tracking results are inconsistent, the weights are adjusted according to their respective performances. For example, if YOLO's accuracy is higher than TransTrack, increase the weight of YOLO.

[0087] 5. Adjust the confidence of YOLO and TransTrack according to the new weights to solve the problem of inconsistent decisions.

[0088] 6. Fusion decision: In the fusion decision stage, the weighted confidence is combined to make the final judgment. For example, if the weighted confidence of YOLO is high, then the judgment of YOLO is more likely to be trusted; if the weighted confidence of TransTrack is high, then the judgment of TransTrack is more likely to be trusted.

[0089] 7. Design a feedback loop to continuously optimize the weighting function and weight adjustment strategy. This can be achieved through machine learning methods such as reinforcement learning, where the system automatically adjusts the strategy based on long-term performance feedback.

[0090] 8. Testing and optimization: Test the effectiveness of the adaptive adjustment method in different traffic scenarios and conditions, and optimize it based on the test results.

[0091] 9. Implementation and monitoring: Implement the adaptive adjustment method into the actual system and continuously monitor its performance to ensure that the system can operate stably and adapt to different traffic scenarios.

[0092] Through the above steps, a method can be designed to adaptively adjust the confidence of the YOLO model and TransTrack model to solve the problem of inconsistent decisions in target recognition and tracking. This method will help improve the accuracy and robustness of the overall system.

[0093] Step 4: Obtain traffic scene indication information by machine vision recognition, including but not limited to: traffic light information and lane information; wherein the traffic light information includes the color information of the current traffic light; the lane information includes but not limited to: the current lane information and the position information of the current vehicle position and the stop line, the lane information includes the left turn lane, the right turn lane, and the execution lane; the position information of the current vehicle position and the stop line includes the vehicle in front of the stop line and the vehicle behind the stop line;

[0094] Integrate traffic scene indication information and tracking data to extract traffic scene indication features and target tracking features; the target tracking features include: features of the tracked target, such as position, size, direction change, etc.;

[0095] Use machine learning or deep learning models to classify target tracking features to identify different types of driving behaviors;

[0096] Based on the identified driving behavior type, determine whether it complies with traffic regulations, including but not limited to running a red light, driving in the wrong lane, driving in the wrong direction, dangerous lane changes, lane departure, speeding, etc., and obtain the behavior analysis results;

[0097] Step 5: Score and warn the driver based on the behavior analysis results; specific operations include:

[0098] Calculate a safety score based on the type and severity of driving behavior;

[0099] When dangerous driving behavior is detected, a warning message is immediately issued; the warning message includes: sound warning, visual warning and tactile warning; among which, the visual warning includes but is not limited to: indicator lights on the dashboard or screen display; among which, the tactile warning includes but is not limited to: seat vibration or seat belt tightening; prompting the driver to take appropriate safety measures. Provide personalized driving improvement suggestions based on the safety score.

[0100] Step 6: Collect the current driving decision information executed by the intelligent connected vehicle algorithm, adjust the parameters of the YOLO model and the TransTrack model based on the current driving decision information, the current traffic scene information and the target tracking data, obtain the updated YOLO model and the TransTrack model, and use the updated YOLO model and the TransTrack model to execute steps 2 to 6.

[0101] Figure 2 A schematic diagram of an application example is shown. The application example takes an intersection with traffic lights (traffic lights) as a scenario, describes the driving behavior of a vehicle when passing through the intersection, and uses the technical solution proposed in this embodiment to judge its driving behavior to determine whether the method complies with conventional traffic regulations during the driving process of the dominant vehicle.

[0102] In a traffic light scenario, the traffic light is first identified by machine vision, and the RGB color is determined based on the recognition frame of the machine vision. The red, green, and yellow lights have obvious RGB color distinction, so this method can be used to determine whether the lane the vehicle is currently in allows forward movement. When the stop line is identified at the same time, it can be determined whether the vehicle continues to drive beyond the stop line when the light is red. However, since the position of the traffic light is relatively mobile for the on-board vision of the moving vehicle, TransTrack is needed for target tracking. In this process, if the vehicle is observed to have exceeded the stop line when the light is red, it is determined to have run a red light.

[0103] In the same scenario, similar to the red light running judgment, whether the vehicle is turning correctly in accordance with the instructions of the turn arrow on the ground can also be identified by the method of this embodiment. There is a common left turn lane with a left turn arrow marked on the ground. The arrow can be easily identified by machine vision. At the same time, taking the traffic light as the reference target, when the traffic light moves to the right in the vehicle's perspective, it is judged as a normal left turn. If the traffic light moves backward, it is executed in the left turn lane, which is a violation of traffic regulations. This process also requires the use of TransTrack's target tracking.

[0104] The method of this embodiment can monitor and evaluate the driver's behavior in real time, provide immediate feedback and improvement suggestions, and thus improve road traffic safety. And through the adaptive learning mechanism, the system can continuously learn from new data to adapt to the ever-changing traffic environment and driving behavior, and achieve continuous improvement in evaluation performance.

[0105] The method of this embodiment can make full use of the information of target detection and target tracking to achieve efficient and accurate estimation of driving behavior.

[0106] Embodiment 2:

[0107] This embodiment discloses a real-time driving behavior evaluation system based on Yolo target detection and TransTrack target tracking technology, which mainly includes:

[0108] A video stream acquisition module is used to acquire road scene video streams in real time;

[0109] The target detection module is used to use the YOLO model to perform target detection on each frame of the road scene video stream to obtain a target detection result 1;

[0110] The target tracking module is used to use the target detection result as the prior feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data; the target tracking data includes: target position change data, target size change data, and target direction change data;

[0111] The driving behavior type recognition module is used to extract the current traffic scene features from each frame image in the road scene video stream using a visual recognition model; and determine the current driving behavior type based on the current traffic scene features and the current target tracking data;

[0112] The driving behavior analysis module is used to determine whether the current driving behavior type complies with traffic regulations and obtain the behavior analysis results;

[0113] The warning module is used to score and warn the current driving behavior based on the behavior analysis results;

[0114] A data storage unit, used for storing captured video streams and behavior analysis results;

[0115] A user interface that allows users to view and export behavior measurement reports;

[0116] A communications module for exchanging data and receiving updates with external systems such as a traffic management center.

[0117] In this embodiment, the target detection result 1 is used as the prior feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data. The specific operations include:

[0118] The TransTrack model performs target detection on each frame of the road scene video stream and obtains target detection result 2;

[0119] A weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model;

[0120] Using the adjusted confidence, the target detection result 1 and the target detection result 2 are fused to determine the final target detection result;

[0121] The spatial relationship is used to associate the final target detection results and the tracking trajectory of the TransTrack model to form target tracking data.

[0122] In this embodiment, a weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model. The specific operations include:

[0123] Determine the evaluation metrics and evaluation metric functions used to measure the performance of the YOLO module and TransTrack model;

[0124] According to the evaluation indicators, the performance data of the YOLO model and TransTrack model are collected in real time;

[0125] According to the following formula, the confidence weight of the YOLO model and the confidence weight of the TransTrack model are dynamically adjusted according to the performance data of the YOLO model and the TransTrack model:

[0126] w YOLO =f(YOLO Performance Metrics)

[0127] w TransTrack =g(TransTrack Performance Metrics)

[0128] Where f(YOLO Performance Metrics) and g(TransTrack Performance Metrics) represent the evaluation index functions of the YOLO model and the TransTrack model respectively; w YOLO Represents the confidence weight of the YOLO model, w TransTrack Represents the confidence weight of the TransTrack model;

[0129] According to the adjusted confidence weights, the confidence of the YOLO model and the confidence of the TransTrack model are determined.

[0130] The system of this embodiment can adaptively learn, optimize according to the driver's behavior pattern, and provide long-term behavior improvement suggestions, thereby realizing efficient and accurate real-time evaluation of driving behavior driven by intelligent driving algorithms, improving road traffic safety during road testing, and being able to continuously optimize according to algorithm-driven driving behavior patterns.

Claims

1. A real-time evaluation method for driving behavior based on YOLO and TransTrack, characterized by: The following steps are involved: Obtain road scene video stream in real time; The YOLO model is used to perform target detection on each frame of the road scene video stream to obtain target detection result 1; The target detection result is used as the prior feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms the target tracking data. The target tracking data includes: target position change data, target size change data and target direction change data; Using the visual recognition model, the current traffic scene features are extracted from each frame of the road scene video stream; According to the current traffic scene characteristics and the current target tracking data, the current driving behavior type is determined; Based on the current driving behavior type, determine whether it complies with traffic regulations and obtain behavior analysis results; Based on the behavior analysis results, current driving behavior is scored and warned; The target detection result is used as a priori feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data. The specific operations include: The TransTrack model performs target detection on each frame of the road scene video stream and obtains target detection result 2; A weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model; Using the adjusted confidence, the target detection result 1 and the target detection result 2 are fused to determine the final target detection result; Using spatial relationships, we associate the final target detection results with the tracking trajectory of the TransTrack model to form target tracking data. The weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model. The specific operations include: Determine the evaluation metrics and evaluation metric functions used to measure the performance of the YOLO model and TransTrack model; According to the evaluation indicators, the performance data of the YOLO model and TransTrack model are collected in real time; According to the following formula, the confidence weight of the YOLO model and the confidence weight of the TransTrack model are dynamically adjusted according to the performance data of the YOLO model and the TransTrack model: w YOLO =f(YOLO Performance Metrics) w TransTrack =g(TransTrack Performance Metrics) Where f(YOLO Performance Metrics) and g(TransTrack Performance Metrics) represent the evaluation index functions of the YOLO model and the TransTrack model respectively; w YOLO Represents the confidence weight of the YOLO model, w TransTrack Represents the confidence weight of the TransTrack model; According to the adjusted confidence weights, the confidence of the YOLO model and the confidence of the TransTrack model are determined.

2. The real-time evaluation method for driving behavior based on YOLO and TransTrack according to claim 1, characterized in that: The evaluation indicators include precision, recall, average precision and tracking accuracy.

3. The real-time driving behavior evaluation method based on YOLO and TransTrack according to claim 1, characterized in that: The warnings include: sound warnings, visual warnings and tactile warnings; the tactile warnings include seat vibrations or seat belt tightening.

4. The real-time driving behavior evaluation method based on YOLO and TransTrack according to claim 1, characterized in that: The following steps are also included: Collect current driving decision information, and adjust the parameters of the YOLO model and the TransTrack model according to the current driving decision information, the current traffic scene information, and the target tracking data to obtain updated YOLO model and TransTrack model; The updated YOLO model and TransTrack model are used to perform target detection and target tracking on each frame image in the road scene video stream.

5. A real-time driving behavior evaluation system based on YOLO and TransTrack, characterized by: include: A video stream acquisition module is used to acquire a road scene video stream in real time; The target detection module is used to use the YOLO model to perform target detection on each frame of the road scene video stream to obtain a target detection result 1; The target tracking module is used to use the target detection result as the prior feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms the target tracking data. The target tracking data includes: target position change data, target size change data and target direction change data; The driving behavior type recognition module is used to extract the current traffic scene features from each frame image in the road scene video stream using a visual recognition model; and determine the current driving behavior type based on the current traffic scene features and the current target tracking data; The driving behavior analysis module is used to determine whether the current driving behavior type complies with traffic regulations and obtain the behavior analysis results; The warning module is used to score and warn the current driving behavior based on the behavior analysis results; The target detection result is used as a priori feature input of the TransTrack model. The TransTrack model combines each frame image in the road scene video stream with the target feature information corresponding to the frame image, performs the target tracking task, and forms target tracking data. The specific operations include: The TransTrack model performs target detection on each frame of the road scene video stream and obtains target detection result 2; A weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model; Using the adjusted confidence, the target detection result 1 and the target detection result 2 are fused to determine the final target detection result; Using spatial relationships, we associate the final target detection results with the tracking trajectory of the TransTrack model to form target tracking data. The weighted adaptive adjustment method is used to adjust the confidence of the YOLO model and the confidence of the TransTrack model. The specific operations include: Determine the evaluation metrics and evaluation metric functions used to measure the performance of the YOLO model and TransTrack model; According to the evaluation indicators, the performance data of the YOLO model and TransTrack model are collected in real time; According to the following formula, the confidence weight of the YOLO model and the confidence weight of the TransTrack model are dynamically adjusted according to the performance data of the YOLO model and the TransTrack model: w YOLO =f(YOLO Performance Metrics) w TransTrack =g(TransTrack Performance Metrics) Where f(YOLO Performance Metrics) and g(TransTrack Performance Metrics) represent the evaluation index functions of the YOLO model and the TransTrack model respectively; w YOLO Represents the confidence weight of the YOLO model, w TransTrack Represents the confidence weight of the TransTrack model; According to the adjusted confidence weights, the confidence of the YOLO model and the confidence of the TransTrack model are determined.

6. The real-time driving behavior evaluation system based on YOLO and TransTrack according to claim 5, characterized in that: Also includes model optimization module; The model optimization module is used to collect current driving decision information, and adjust the parameters of the YOLO model and the TransTrack model according to the current driving decision information, the current traffic scene information and the target tracking data to obtain the updated YOLO model and the TransTrack model.

Images