Path Planning Method Combining Small-Scale Traffic Sign Recognition and DWA Algorithm

Abstract

The present invention relates to the technical fields of image recognition and navigation, and discloses a path planning method combining small-scale traffic sign recognition and the DWA algorithm. Road surface images are collected and input into a pre-trained offline small-scale traffic sign detection model to detect the positions, labels, and weights of traffic signs in real time. Then, the obtained labels and weights are used as the trigger conditions for improving the DWA algorithm. After the trigger conditions are met, local path planning is performed through the improved DWA algorithm. The small-scale traffic sign detection model is based on the YOLOv5s network, with an FEM module and an ECA-SAM parallel attention mechanism introduced in the Neck, and a small-scale detection head added to the head. The improved DWA algorithm mainly adds algorithm trigger conditions and improves the evaluation function. The present invention realizes vehicle obstacle avoidance during the process of autonomous driving based on traffic signs, avoids the DWA algorithm from always performing obstacle avoidance planning, and improves the adaptability and safety of the DWA algorithm in complex environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition and navigation, and specifically relates to a path planning method that combines small-scale traffic sign recognition and the DWA algorithm. Background Art

[0002] The research and application of autonomous driving technology have become an important direction in the modern automotive industry. Autonomous driving technology can not only improve traffic efficiency, reduce traffic accidents, but also relieve the burden on drivers. However, the application of autonomous driving technology in complex environments still faces many challenges, such as the accurate recognition of traffic signs, path planning in complex environments, real-time obstacle avoidance, etc.

[0003] Small-scale traffic signs refer to traffic sign images that can be captured (such as traffic lights, left-turn signs, etc.) being too small. In the prior art, the YOLOv5s object detection algorithm has certain advantages in traffic sign recognition, but there are still problems of false detection and missed detection in the process of small-scale traffic sign detection. The Dynamic Window Approach (DWA) is a local path planning algorithm for obstacle avoidance, which generates possible speed sampling points by combining data on speed and acceleration. The algorithm can generate possible motion trajectories, and then evaluate these trajectories based on factors such as distance and collision risk, and select the optimal one from the candidate trajectories as the path of the robot. However, due to the defects of the evaluation function of the DWA algorithm itself, there are problems such as the path being close to obstacles and trajectory oscillation in path planning. In addition, existing DWA algorithms and improvements to the DWA algorithm mainly consider simply improving the evaluation function of DWA to achieve the real-time performance and accuracy of the path planned by the DWA algorithm. For example, in the improvement part of the DWA algorithm in the invention with the publication number CN 118938924 A, "A Dynamic Obstacle Avoidance Method Based on the Fusion of Kalman Filter and Improved DWA Algorithm", the improved evaluation function is used to score the speed pairs in the speed window of the DWA algorithm, and the speed pair with the highest score is selected until the target point is found to solve the problem of being easily trapped in the dynamic local optimum. However, in the actual road surface environment, in addition to considering factors such as obstacles and speed, the DWA algorithm also needs to consider the information of traffic signs on the road surface. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a path planning method that combines small-scale traffic sign recognition and the DWA algorithm, so as to combine the recognition of traffic signs and the DWA algorithm for local obstacle avoidance, thereby achieving more efficient and safer autonomous driving.

[0005] To solve the above technical problems, the present invention provides a path planning method combining small-scale traffic sign recognition and the DWA algorithm, including: an autonomous driving unmanned vehicle collects road surface images, and after uniformly scaling the size, inputs them into a pre-trained offline small-scale traffic sign detection model to detect the position, label, and weight information of traffic signs in real time. Then, the obtained label and weight information are used as the trigger conditions for improving the DWA algorithm. After meeting the trigger conditions, local path planning is performed through the improved DWA algorithm.

[0006] The small-scale traffic sign detection model is improved based on the YOLOv5s network. First, cross-level cascading is added, and the feature maps extracted by the C3 module in the Backbone are concatenated with the feature maps extracted by two Conv modules in the Neck in the channel dimension. Secondly, a feature enhancement module is added to each layer of the downsampled feature pyramid in the Neck of the YOLOv5s network, and the ECA-SAM parallel attention mechanism is added between every two layers of the upsampled feature pyramid. Additionally, a small-scale feature detection network is added in the Neck. Thirdly, a small-scale detection head is added to the head of the YOLOv5s network.

[0007] The improved DWA algorithm includes adding an algorithm trigger condition based on the detection results of the small-scale traffic sign detection model and improving the evaluation function for evaluating predicted trajectories.

[0008] As an improvement to the path planning method of the present invention combining small-scale traffic sign recognition and the DWA algorithm:

[0009] The structure of the small-scale feature detection network is the same as that of the original feature detection network of the YOLOv5s network and they are arranged in parallel. The structure of the small-scale detection head is the same as that of the original three detection heads of the YOLOv5s network and they are arranged in parallel. The small-scale feature detection network is connected to the small-scale detection head.

[0010] As a further improvement to the path planning method of the present invention combining small-scale traffic sign recognition and the DWA algorithm:

[0011] The offline training process of the small-scale traffic sign detection model is as follows:

[0012] The training set, validation set, and test set in the Bosch Small Traffic Signs dataset are used for offline training. Each sample image size in the dataset is uniformly scaled and then data augmentation operations are performed, including image translation, image scaling, and random rotation. Training parameters are set, and the training set images and labels are input into the small-scale traffic sign detection model. The loss function is calculated and the gradient is obtained through backpropagation until the loss function on the training set and validation set no longer decreases. Then, the test dataset is input into the trained model to verify the recall rate and accuracy.

[0013] As a further improvement of the path planning method combining small-scale traffic sign recognition and the DWA algorithm of the present invention:

[0014] The algorithm triggering condition of the improved DWA algorithm is:

[0015] (4)

[0016] Wherein, represents the weight of the traffic sign detected by the small-scale traffic sign detection model, represents the weight triggering threshold.

[0017] As a further improvement of the path planning method combining small-scale traffic sign recognition and the DWA algorithm of the present invention:

[0018] The evaluation function for evaluating the predicted trajectory of the improved DWA algorithm is:

[0019] (13)

[0020] Wherein, is the weight coefficient, is a fixed constant; is the normalization coefficient; represents the angle deviation between the trajectory end orientation and the target point; represents the minimum distance from the improved trajectory end to the obstacle; represents the speed; is the safety radius.

[0021] As a further improvement of the path planning method combining small-scale traffic sign recognition and the DWA algorithm of the present invention:

[0022] The minimum distance from the improved trajectory end to the obstacle is:

[0023] (11)

[0024] The value range is:

[0025] (12).

[0026] The beneficial effects of the present invention are mainly reflected in:

[0027] 1. The DWA path planning algorithm of the present invention is integrated with a small-scale traffic sign detection model. The recognized traffic sign information is input into the DWA algorithm for local obstacle avoidance, realizing vehicle obstacle avoidance during the automatic driving process based on traffic signs, avoiding the continuous obstacle avoidance planning of the DWA algorithm during vehicle driving, and improving the adaptability and safety of the DWA algorithm in complex environments.

[0028] 2. The small-scale traffic sign detection model of the present invention introduces a parallel attention mechanism based on the YOLOv5s network, significantly improving the detection accuracy of small-scale traffic signs and solving the problem of missed detection in the prior art.

[0029] 3. The present invention optimizes the path planning performance. By improving the evaluation function of the DWA algorithm, it avoids the problem of the path being close to obstacles, reduces trajectory oscillation, and improves the efficiency and safety of path planning.

[0030] 4. Through the path planning method combining small-scale traffic sign recognition of the present invention, it is possible to reduce costs and improve cost performance, reduce the dependence on expensive hardware, maintain high performance at the same time, and improve the cost performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The following further elaborates on the specific implementation manners of the present invention with reference to the drawings.

[0032] Figure 1 It is a schematic diagram of the process of constructing and offline training the small-scale traffic sign detection model of the present invention;

[0033] Figure 2 It is a schematic diagram of the structure of the small-scale traffic sign detection model of the present invention;

[0034] Figure 3 For Figure 2 It is a schematic diagram of the structure of the ECA-SAM parallel attention module in

[0035] Figure 4 It is a schematic diagram of adding a small-scale detection head and cross-level cascading in the small-scale traffic sign detection model of the present invention;

[0036] Figure 5 It is a schematic diagram of the detection result of the small-scale traffic sign detection model of the present invention;

[0037] Figure 6 It is a simulation effect diagram of the path planning of the traditional DWA algorithm;

[0038] Figure 7 It is a simulation effect diagram of the path planning of the improved DWA algorithm of the present invention;

[0039] Figure 8It is the path planning speed curve graph of the traditional DWA algorithm;

[0040] Figure 9 It is the path planning speed curve graph of the improved DWA algorithm of the present invention. Specific implementation manners

[0041] The present invention will be further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto:

[0042] Embodiment 1. A path planning method combining small-scale traffic sign recognition and the DWA algorithm, which is used for autonomous driving adapting to complex traffic environments. The autonomous driving unmanned vehicle obtains small-scale traffic sign information in real time through a pre-trained offline small-scale traffic sign detection model, and at the same time obtains obstacle information on the driving path through its own radar. Combining vehicle speed information (maximum linear speed, maximum angular speed, acceleration, angular acceleration, linear speed resolution, angular speed resolution), the dynamic window method (DWA) using an improved evaluation function is used for local path planning, making the driving more real-time.

[0043] Step 1. A small-scale traffic sign detection method based on the improved YOLOv5s network

[0044] The present invention first constructs the small-scale traffic sign detection model of the present invention based on the improved YOLOv5s network, and then uses the Bosch Small Traffic Signs dataset to perform offline training on the small-scale traffic sign detection model, so as to obtain a small-scale traffic sign detection method that can be used online. The process of constructing and offline training the small-scale traffic sign detection model is as Figure 1 shown, specifically:

[0045] Step 1.1: Obtain an offline training dataset containing traffic signs in the natural environment, and perform data preprocessing and data augmentation.

[0046] The present invention uses the Bosch Small Traffic Signs dataset as the offline training dataset of the small-scale traffic sign detection model. The numbers of training samples, test samples and validation samples in the Bosch Small Traffic Signs dataset are 3500, 1000 and 500 respectively.

[0047] The data preprocessing steps include:

[0048] (1)Image size unification: Since the sizes and resolutions of the images in the dataset are inconsistent, all images need to be adjusted to a unified size for easy model processing. Resize the images in the Bosch Small Traffic Signs dataset to 640×1280 while maintaining the aspect ratio of the original images, reducing the computational load of the model and accelerating the forward propagation speed of the model;

[0049] (2)Establishing the relationship between categories and IDs: Establish a unique ID for each category in the dataset for convenient subsequent model training and evaluation.

[0050] The data augmentation steps include:

[0051] (1)Translation: Randomly translate the traffic signs in the image to simulate traffic signs appearing in different positions.

[0052] (2)Scaling: Randomly scale the image to simulate the size changes of traffic signs at different distances.

[0053] (3)Random rotation: Rotate the image by a random angle to enhance the model's ability to recognize traffic signs in different directions.

[0054] Convert the format of the label file to the txt format required by YOLOv5s. Each image corresponds to a label file, and each line of the file represents the information of a target. The information includes the corresponding list name, the center coordinates, and the length and width of the bounding box.

[0055] Step 1.2: Build a small-scale traffic sign detection model based on the improved YOLOv5s;

[0056] Step 1.2.1: In the backbone network, the input image is first sliced and stitched through the Focus module, and then passes through a series of convolutional layers (Conv + C3) and the Spatial Pyramid Pooling (SPP) layer to extract features. In the neck network, a Feature Enhancement Module (FEM) is added to each layer of the downsampled feature pyramid. The FEM module can effectively solve the problem of information loss caused by using dilated convolutions with a fixed dilation rate alone by combining multiple dilated convolutions with different dilation rates, thus optimizing the processing effect.

[0057] Then, the ECA-SAM parallel attention mechanism is added between every two layers of the upsampled feature pyramid. Since the original network model faces many challenges in solving small-scale traffic signs, especially the lack in accurately allocating and processing target region features, the ECA-SAM parallel attention mechanism is introduced into the feature fusion layer of the Neck part of the original YOLOv5s. Specifically, the process is as follows Figure 2As shown, the feature map first undergoes fine processing by the attention module and then is passed bottom-up, thus enhancing the network's ability to learn and recognize the features of small-scale targets.

[0058] Finally, multiple object detection heads (Detect) in the head network (Head) are used to predict the bounding boxes and classes of objects. The object detection head is the part used to perform object detection on the feature pyramid, and it includes convolutional layers, pooling layers, fully connected layers, etc. In the original YOLOv5s model, the detection head module is mainly responsible for performing multi-scale object detection on the feature map extracted by the backbone network, including large, medium, and small detection heads. The small-scale traffic sign detection model of the present invention adds a small-scale detection head at the topmost layer to achieve small-scale object detection. The entire Head part network realizes the fusion of features at different levels through the concatenation (Concat) of feature maps, thereby improving the accuracy of object detection.

[0059] The ECA-SAM parallel attention mechanism is as Figure 3 shown. The ECA-SAM parallel attention mechanism integrates the channel attention (ECA) module and the spatial attention (SAM) module, performs a binary process on the input, and forms two parallel paths. The upper part focuses on channel attention and directly learns the features extracted by the global average pooling layer through the fine processing of the ECA module. The lower part generates the weight values of spatial attention through the SAM module. The weights of the outputs of the two parts are respectively normalized. Then, the normalized weights are fused using a dot product operation to output the weights of the parallel attention mechanism. The weights of the ECA-SAM parallel attention mechanism are shown in the following formula.

[0060] (1)

[0061] In the formula, represents the final feature output by the ECA-SAM parallel attention mechanism; represents the sigmoid function; represents the weight output by the channel attention; represents the element-wise multiplication operation with the original input feature, represents the weight value of the spatial attention output.

[0062] The small-scale traffic sign detection model of the present invention uses YOLOv5s as the basic network architecture, introduces the ECA-SAM parallel attention mechanism by integrating channel attention and spatial attention, and then further judges and recognizes the images collected by the high-speed camera through weights to exclude the interference of non-target objects in the complex traffic environment.

[0063] Step 1.2.2: Based on the three detection heads of YOLOv5s, a new detection head is added to the head of YOLOv5s as a small-scale detection head. As shown in Figure 4 , the newly added small-scale detection head (Detect4) has the same structure as the original three detection heads (Detect1, Detect2, and Detect3) of the YOLOv5s network and presents a parallel structure. At the same time, a small-scale feature detection network is added to the neck. The added small-scale feature detection network has the same structure as the original feature detection network of YOLOv5s and is parallel to it. The small-scale feature detection network is connected to the small-scale detection head (Detect4) to construct a four-channel detection structure. The newly added detection part is shown in Figure 4 the green box part. This measure helps to detect small-scale targets, thereby further improving the network's recognition ability for medium and long-distance targets. In addition, cross-level cascading is added to the network to ensure the effective transmission of shallow feature information of the network backbone to the feature fusion layer. As shown by the red line in Figure 4 , the feature map extracted by the C3 module in the Backbone is concatenated with the feature maps of the same width and height extracted by the two Conv modules in the Neck part in the channel dimension. This measure aims to strengthen the backpropagation mechanism of the gradient, prevent the occurrence of gradient attenuation, and thus significantly reduce the loss of feature information that may occur when identifying small-scale traffic signs.

[0064] Step 1.3: Input the training images and labels into the built detection model for training until convergence;

[0065] Step 1.3.1: Set the parameters of the small-scale traffic sign detection model: the batch size is 32, the initial learning rate r is 1e-5. If the change of the loss curve is slow during training, the learning rate is increased.

[0066] Step 1.3.2: Input the training set images and labels into the improved YOLOv5s network (small-scale traffic sign detection model), calculate the loss function and backpropagate to calculate the gradient until the loss functions (classification loss function Binary Cross Entropy, abbreviated as BCE, localization loss function Complete Intersection over Union, abbreviated as CIoU, and weight loss function BCE) on the training set and validation set no longer decrease, and at the same time the output evaluation indicators (accuracy and recall) no longer increase. The model parameters at this time are the trained model parameters. The following are the calculation formulas of the BCE loss function and the CIoU loss function.

[0067] (2)

[0068] Among them, is a binary label 0 or 1, is the probability that the output belongs to the label, represents the number of groups of objects predicted by the model.

[0069] (3)

[0070] Among them, is the intersection over union ratio, is the distance between the centers of the predicted box and the ground truth box, is the diagonal distance of the minimum bounding rectangle, is the weight factor, is the correction factor.

[0071] Then, the test set images are input into the trained model. The recall rate of the test data set reaches 88.7%, and the accuracy rate reaches 91.6%, thus completing the offline training to obtain a small-scale traffic sign detection model that can be used online.

[0072] Step 1.4: Online use;

[0073] When the autonomous driving unmanned vehicle is driving, images are automatically collected through the high-speed camera carried on the vehicle body, and then the images are input into the small-scale traffic sign detection model trained offline. The model detects the position, label, and weight information of the traffic signs, as shown in the example Figure 5 shown, for the next step of path planning.

[0074] Step 2: Improve the dynamic window approach (DWA) of the evaluation function for local path planning

[0075] During the driving process of the autonomous driving unmanned vehicle, the local grid map is obtained through the on-vehicle radar to obtain obstacle information, providing map information for the local planning of the DWA algorithm. At the same time, after successfully identifying traffic signs (including the position, label, and weight information of traffic signs) through Step 1.4, the label and weight information of the traffic signs are output, and the obtained label and weight information are used as the input signal for starting the DWA algorithm in the present invention for motion control. Only when those traffic signs are detected and the weights meet the DWA trigger conditions, the DWA algorithm will assist in motion control, avoiding situations where the DWA algorithm has been performing obstacle avoidance planning when the vehicle is going straight, improving the adaptability and safety of the DWA algorithm.

[0076] The traditional DWA algorithm obtains multiple sets of speed combinations by sampling the dynamic speed window , and evaluates the predicted trajectories of different speed combinations within a certain time through an evaluation function, and selects the speed combination corresponding to the trajectory with the highest evaluation as the motion speed of the unmanned vehicle. The traditional DWA algorithm can be divided into map building, determining the speed window, predicting future trajectories, and evaluating predicted trajectories.

[0077] The improved DWA algorithm of the present invention adds an algorithm trigger condition based on the detection results of the small-scale traffic sign detection model and an improvement to the evaluation function for evaluating the predicted trajectory. It can be divided into algorithm trigger, map establishment, determination of the speed window, prediction of the future trajectory, and evaluation of the predicted trajectory using the improved evaluation function, as follows:

[0078] (1) The algorithm trigger needs to meet:

[0079] (4)

[0080] Among them, represents the weight corresponding to the traffic sign detected by the small-scale traffic sign detection model, represents the set weight trigger threshold.

[0081] (2) Map establishment:

[0082] The millimeter-wave radar scans to obtain the obstacle information set , and obtains the map required for DWA path planning through the obstacle information set . The subsequent DWA algorithm needs to perform path planning in the map .

[0083] (3) The DWA dynamic speed window is:

[0084] (5)

[0085] Among them,

[0086]

[0087] (6)

[0088] (7)

[0089] Among them, , are the maximum linear speed and maximum angular speed of the unmanned vehicle respectively; , are the minimum linear speed and minimum angular speed of the unmanned vehicle respectively; , are the current linear speed and angular speed of the unmanned vehicle respectively; , are the maximum deceleration of the unmanned vehicle; , are the maximum acceleration of the unmanned vehicle, is the sampling interval, represents the minimum distance from the end of the trajectory to the obstacle.

[0090] ​(4)The prediction of the DWA trajectory specifically includes:

[0091] Uniformly sample in the DWA dynamic speed window, and the number of optional speed combinations is: For:

[0092] (8)

[0093] Among them, is the linear velocity resolution, is the angular velocity resolution.

[0094] Select the predicted trajectory of the speed combination as:

[0095] (9)

[0096] Among them is the pose of the unmanned vehicle at time k, is the sampling interval.

[0097] (5)Evaluate the predicted trajectory

[0098] Since the DWA algorithm does not consider the directionality, motion smoothness, and rapidity of reaching the target point of the trajectory to be evaluated. Therefore, an evaluation function can be added to optimize the DWA algorithm to improve the stability of the robot's motion direction, the smoothness of motion, and reduce the time for the robot to reach the end point. The role of the evaluation function is to evaluate the quality of multiple groups of predicted trajectories obtained by sampling, determine the optimal trajectory, and select the corresponding speed combination as the motion speed of the robot. The evaluation function of the traditional DWA algorithm is shown in the following formula:

[0099] (10)

[0100] Among them, represents the angle deviation between the end of the trajectory and the target point. The smaller the deviation, the higher the evaluation; represents the minimum distance from the end of the trajectory to the obstacle. The larger the value, the farther away from the obstacle, and the higher the evaluation; represents the speed. The faster the speed, the higher the evaluation; is the weight coefficient, is the normalization coefficient.

[0101] In the traditional DWA algorithm is defined as the minimum value of the connection line between the end of all predicted trajectories and the center of the obstacle. Considering the size of the obstacle and the size of the robot, the safety radius is defined as The new distance evaluation is defined as:

[0102] (11)

[0103] When there are no obstacles around the vehicle body, to prevent excessive influence, the value range can be defined as:

[0104] (12)

[0105] Regarding the problem of path oscillation when the vehicle body enters a dense obstacle area, the DWA evaluation function is improved: reducing the attention to the target point and compensating the speed at the same time. Define the obstacle distance factor and the new evaluation function can be obtained as:

[0106] (13)

[0107] Among them, is a fixed constant.

[0108] Finally, each predicted trajectory of the improved DWA algorithm is evaluated through the new evaluation function to output the evaluation score, and the trajectory with the highest evaluation score is selected as the best trajectory.

[0109] Experiment:

[0110] 1. Verify the feasibility of the improved DWA algorithm during the movement through simulation.

[0111] Table 1 shows the motion parameters of the simulated unmanned vehicle using the improved DWA algorithm of the present invention, and Table 2 shows the DWA algorithm parameters. The simulation operation effect of using the improved DWA algorithm of the present invention is as Figure 7 shown, and the speed curve is as Figure 9 shown.

[0112] Table 1 Motion parameters of the simulated unmanned vehicle

[0113] Maximum linear velocity (m / s) Maximum angular velocity (rad / s) <![CDATA[Acceleration (m / s 2 )]]> <![CDATA[Angular acceleration (rad / s 2 )]]> Linear velocity resolution (m / s) Angular velocity resolution (rad / s) 3 0.75 0.3 1.745 0.03 0.175

[0114] Table 2 DWA algorithm parameters

[0115]

[0116] The simulation operation effect of the traditional DWA algorithm is as Figure 6 shown, and the speed curve is as Figure 8 shown. The traditional DWA algorithm shows phenomena such as trajectory oscillation, speed curve fluctuation, and even vehicle collision with obstacles in a dense obstacle environment; while from Figure 7 and Figure 9 it can be seen that the path planned by the DWA algorithm with the improved evaluation function is smooth and without oscillation, the speed curve is overall stable and without fluctuation, which better meets the kinematic requirements of the robot in the actual environment, and the overall planning speed of the improved DWA algorithm is 4.7% higher than that of the traditional DWA algorithm.

[0117] 2. Experiments on Small-Scale Object Detection Models

[0118] Ablation experiments were conducted on the ECA-SAM parallel attention mechanism and the multi-detection head network respectively. In the experiments on the ECA-SAM parallel attention mechanism, not only the original YOLOv5s network was compared, but also the YOLOv5s network using the channel attention mechanism ECANet and the YOLOv5s network using the spatial attention mechanism SAM were compared.

[0119] Then, in the comparative experiment for model performance evaluation, the divided test samples were input into the small-scale traffic sign detection model, and the precision and recall rates of the model for traffic sign detection were output. These were compared with the evaluation indicators of the original YOLOv5s network to prove the effectiveness of the improved algorithm.

[0120] 2.1 Ablation Experiments:

[0121] Comparative experiments were conducted before and after introducing different attention mechanisms and adding detection heads. The experimental results are shown in Tables 3 and 4.

[0122] Table 3 Comparative Experiments after Introducing Different Attention Mechanisms

[0123] Network Recall (%) Precision (%) YOLOv5s 86.4 90.5 YOLOv5s+ECAnet 86.1 90.7 YOLOv5s+SAM 86.6 90.8 YOLOv5s+ECA-SAM 87.5 91.3

[0124] As can be seen from Table 3, after introducing three different attention mechanisms into the YOLOv5s network, the detection accuracy of the model has been improved. Among them, the model incorporating the ECA-SAM parallel attention mechanism performs the best, with its detection accuracy increased by 0.8% compared to the original model. For ECANet and SAM, although the improvement in network performance is not significant, with only a 0.2% and 0.3% increase in detection accuracy respectively, these small improvements are still worthy of recognition. These results indicate that by reasonably introducing attention mechanisms into the YOLOv5s network, its performance in object detection tasks can be significantly improved. In summary, the ECA-SAM parallel attention mechanism has certain advantages compared with common attention mechanisms.

[0125] Table 4 Comparative Experiments on Adding Detection Heads

[0126] Network Recall (%) Precision (%) YOLOv5s 86.4 90.5 YOLOv5s+Detect 86.5 90.7

[0127] As can be seen from Table 4, after adding an additional small-scale detection head to the YOLOv5s network, the detection accuracy of the model has increased by 0.2%.

[0128] 2.2 Comparative Experiments

[0129] The results of the comparative experiment with YOLOv5 are shown in Table 5.

[0130] Table 5 Comparative experiments with YOLOv5s

[0131] Network Recall (%) Precision (%) YOLOv5s 86.4 90.5 YOLOv5s+ECA-SAM 87.5 91.3 YOLOv5s+ECA-SAM+Detect(ours) 88.7 91.6

[0132] As shown in Table 5, introducing the ECA-SAM parallel attention mechanism into the YOLOv5s network can improve the performance of the network. Compared with the original YOLOv5s algorithm, the detection accuracy and recall of the traffic sign detection algorithm we proposed are improved by 1.1% and 2.3% respectively. Compared with the YOLOv5s algorithm that only introduces the ECA-SAM attention mechanism, the detection accuracy and recall are improved by 0.3% and 1.2% respectively. This fully demonstrates the advantages of the algorithm in traffic sign detection, especially in small-scale traffic sign detection.

[0133] Finally, it should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A path planning method combining small-scale traffic sign recognition and DWA algorithm, characterized by: The autonomous driving vehicle collects road images, uniformly scales them, and then inputs them into the offline trained small-scale traffic sign detection model to detect the location, label, and weight information of traffic signs in real time. The obtained label and weight information are then used as the trigger conditions of the improved DWA algorithm. Only when those traffic signs are detected and the weights meet the DWA trigger conditions, the DWA algorithm will assist in motion control. After the trigger conditions are met, the improved DWA algorithm will perform local path planning. The small-scale traffic sign detection model is improved based on the YOLOv5s network. First, a cross-layer cascade is added to concatenate the feature map extracted by the C3 module in Backbone and the feature map extracted by the two Conv modules in Neck in the channel dimension; secondly, a feature enhancement module is added to each layer of the down-sampling feature pyramid of the YOLOv5s network Neck, and then the ECA-SAM parallel attention mechanism is added between every two layers of the up-sampling feature pyramid, and another small-scale feature detection network is added in Neck; Again, add a small-scale detection head to the YOLOv5s network head; The improved DWA algorithm includes adding an algorithm triggering condition based on the detection results of a small-scale traffic sign detection model and improving the evaluation function for evaluating the predicted trajectory; The algorithm triggering conditions of the improved DWA algorithm are: P sign >P set ; Among them, P sign represents the weight of the traffic sign detected by the small-scale traffic sign detection model, P set Represents the weight trigger threshold.

2. The path planning method combining small-scale traffic sign recognition and DWA algorithm according to claim 1, characterized in that: The small-scale feature detection network has the same structure as the original feature detection network of the YOLOv5s network and is arranged in parallel. The small-scale detection head has the same structure as the original three detection heads of the YOLOv5s network and is arranged in parallel. The small-scale feature detection network is connected to the small-scale detection head.

3. The path planning method combining small-scale traffic sign recognition and DWA algorithm according to claim 2, characterized in that: The offline training process of the small-scale traffic sign detection model is as follows: The training set, validation set and test set in the Bosch Small Traffic Signs dataset are used for offline training. The size of each sample image in the dataset is uniformly scaled before data augmentation operations, including image translation, image scaling and random rotation. The training parameters are set, and the training set images and labels are input into the small-scale traffic sign detection model. The loss function is calculated and the gradient is back-propagated until the loss function on the training set and validation set no longer decreases. Then the test dataset is input into the trained model to verify the recall and accuracy.

4. The path planning method combining small-scale traffic sign recognition and DWA algorithm according to claim 3, characterized in that: The ECA-SAM parallel attention mechanism binary processes the input by integrating the channel attention (ECA) module and the spatial attention (SAM) module to form two parallel paths; The upper part focuses on channel attention. Through the fine processing of the ECA module, it realizes the direct learning of the features extracted by the global average pooling layer. The lower part generates the weight value of the spatial attention through the SAM module. The weights of the two parts are normalized respectively, and then the normalized weights are fused by point multiplication operation to output the weight of the parallel attention mechanism. The weight of the ECA-SAM parallel attention mechanism is shown in the following formula; In the formula, F ECA-SAM represents the final feature output by the ECA-SAM parallel attention mechanism; σ represents the sigmoid function; F ECANet Represents the weight of the channel attention output; Represents the element-by-element multiplication operation with the original input feature, F SAM Represents the weight value of the spatial attention output.

5. The path planning method combining small-scale traffic sign recognition and DWA algorithm according to claim 4, characterized in that: The evaluation function G(v, w) of the evaluation prediction trajectory of the improved DWA algorithm is: Among them, α, β, γ are weight coefficients, a∈(0,1) is a fixed constant; σ is a normalization coefficient; heading(v,w) represents the angular deviation between the end of the trajectory and the target point; dist * (v, w) represents the minimum distance from the end of the improved trajectory to the obstacle; vel(v, w) represents the speed; R is the safety radius.

6. The path planning method combining small-scale traffic sign recognition and DWA algorithm according to claim 5, characterized in that: The minimum distance from the end of the improved trajectory to the obstacle is: dist * (v,w)=dist(v,w)-R; The value range is:

Images