Large vehicle right blind area pedestrian anti-collision detection method based on improved YOLOv8 model

Abstract

The invention discloses a large vehicle right blind area pedestrian anti-collision detection method based on an improved YOLOv8 model, a feature recombination up-sampling module in the model introduces a double-layer routing attention mechanism and a lightweight up-sampling operator, and the double-layer routing attention mechanism filters out non-critical value pairs in a coarse region of a feature map; the capability of extracting small target features in the feature map is improved; the lightweight upsampling operator performs block processing on an input feature map, performs feature recombination on each block by using convolution operation, and enhances a sensing range so that the upsampled feature map can be fused with more context information. The improved YOLOv8 model is tested and trained, so that the average accuracy of a detection result can be improved. A video set standard method is adopted to replace a manual method, and a photoelectric and video combined method is adopted to judge the response time, so that the response speed when an alarm occurs is improved, and traffic safety accidents are reduced.

Description

Technical Field

[0001] The present invention relates to the field of vehicle application technology, and in particular to a pedestrian anti-collision detection method in the right blind spot of a large vehicle based on an improved YOLOv8 model. Background Art

[0002] The cab of engineering vehicles is high, the vehicle structure is special, and the inner and outer wheels are different. The driver has blind spots when observing pedestrians while driving. At the same time, due to the heavy load, large turning radius, long transportation time, and many night working hours of engineering vehicles, if you are not careful when driving, it is easy to cause accidents such as pedestrians and electric vehicles being destroyed and killed. According to statistics from the traffic management department, more than 70% of engineering vehicle accidents are caused by pedestrian blind spots. Therefore, companies across the country have developed blind spot pedestrian recognition systems. The essence of the system is to collect monitoring images of the right blind spot of the engineering vehicle through a camera, judge the movement of pedestrians through image recognition technology, and then judge whether there is a possible danger. If there is, the system will immediately issue an alarm to remind the driver to take appropriate measures to reduce casualties. It can be seen that how to judge the movement of pedestrians is the top priority of the system. There is no detailed and effective testing method in the existing national standards GB / T 19056-2021, JT / T794-2019, JT / T 1076-2016, JT / T 808-2019, and JT / T 1078-2016. Due to the current limitations of testing methods, it is impossible to provide testing manufacturers with unified time, unified scenes, unified routes, and unified warning conditions. The detection plan still remains in the manual detection method, which is highly subjective, inefficient, and inconsistent. Taking the same model product as an example, the manual test comparison data is shown in Table 1.

[0003] Table 1 Comparison data of manual detection

[0004] The purpose of this scheme is to provide a set of standardized police video collection standards, provide the standards to the factories, build a unified test platform for them, and quickly determine the police situation according to the recognition results given by the participating manufacturers, and improve the accuracy of police recognition, the false alarm rate of police recognition, and the missed alarm rate of police recognition. Since the police situation is affected by many factors such as the pedestrian's clothing, height, weight, whether he holds an umbrella, whether he rides a bicycle, whether he wears a helmet, whether he rides with others, the brightness and shadow of the scene where the pedestrian is located, and the weather factors such as rain, snow, and fog, the police situation scene becomes extremely complex. If a complete video collection standard is to be made, it may take a long time and continuously to collect various police scenes, which is unrealistic in actual work. According to statistics from the traffic control department, the proportion of serious casualties in engineering vehicles when they just start to turn is the highest, accounting for about 80%. Therefore, this scheme considers first recording the pedestrian police video collection standard in the static state of the vehicle, and then supplementing and improving the dynamic pedestrian police video collection standard based on practical experience. The video set standard is detected through the YOLOv8 model. The Transformer, with the advantage of the self-attention mechanism, can capture the contextual feature information of the image. It is widely used in target detection models to solve the problem of small targets not being detected due to little feature information. However, the conventional Transformer model has a large number of parameters. If the model is directly used to operate on the feature map globally, it will increase the amount of calculation and occupy a lot of memory, thus affecting the detection reasoning speed, and will also waste a lot of time processing non-target areas in the image. In addition, the upsampling in the YOLOv8 model mostly uses the nearest neighbor linear interpolation method, which will cause the texture details in the image to be lost and cannot accurately restore the details of the original data. Summary of the invention

[0005] The present invention provides a method for detecting pedestrians in the right blind spot of a large vehicle based on an improved YOLOv8 model, so as to solve the technical problems raised in the background technology.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] A method for detecting pedestrians in the right blind spot of a large vehicle based on an improved YOLOv8 model includes the following steps: S1. Data preprocessing: S11. Collect images of pedestrian warning situations as a data set. Pedestrian warning situations include daytime positive pedestrian warning situations, daytime reverse pedestrian warning situations, nighttime positive pedestrian warning situations, and nighttime reverse pedestrian warning situations. Each type of warning situation collects n images, where n≥1600. S12, standardize and enhance the image data; S13, annotating the target area of ​​the image to obtain annotation data, and dividing the image into a training set and a test set; the ratio of the training set to the test set is 4:1; S2. Model training and application: S21. Use the test set to test the improved YOLOv8 model, and save the model parameters when the average recognition accuracy is the highest, that is, the optimal model parameters; The improved YOLOv8 model includes: backbone network module, feature reorganization upsampling module and neck network module; Among them, the feature reorganization upsampling module introduces a double-layer routing attention mechanism and a lightweight upsampling operator. The SPPF module of the YOLOv8 model is used to sample the image to obtain a feature map. The BRA double-layer routing attention mechanism filters irrelevant key value pairs in the coarse area of ​​the feature map, and then divides the blocks into blocks through a lightweight upsampling operator, and uses a convolution operation to reorganize the features of each block. S22, load the optimal model parameters into the improved YOLOv8 model, and train the model using the labeled data; S23. Apply the trained model to the actual image for detection, and record the response time and detection results.

[0008] As a further optimization of the present invention, in step S11, the four types of police situations all include static independent pedestrian scenes, umbrella pedestrian scenes, bicycle pedestrian scenes, raincoat bicycle pedestrian scenes, raincoat electric vehicle pedestrian scenes and cargo electric vehicle pedestrian scenes, and the personnel, clothing, venues, light intensity, etc. in each scene are completely unified.

[0009] As a further preferred embodiment of the present invention, in step S12, the image is scaled and cropped to meet the input requirements of the improved YOLOv8 model, and then the random rotation, flipping, and brightness adjustment methods are used to increase the diversity of image data and enhance the robustness of the model.

[0010] As a further preferred embodiment of the present invention, the specific steps of step S21 are: The improved YOLOv8 model is tested using the test set to calculate precision, recall, and average accuracy. Precision refers to the percentage of samples predicted as positive examples among all positive examples, and recall refers to the percentage of samples predicted as positive examples among all predicted examples. The calculation formulas for precision and recall are as follows: , Among them, TP refers to the number of samples correctly predicted by the model as positive, FP refers to the number of samples incorrectly predicted by the model as positive, TN refers to the number of samples correctly predicted by the model as negative, and FN refers to the number of samples incorrectly predicted by the model as negative. With recall as the horizontal axis and precision as the vertical axis, a PR curve is formed as the threshold changes. The area enclosed by the PR curve and the horizontal and vertical axes is the average accuracy. The formula is as follows: , The corresponding average accuracy of the detection index is obtained by averaging multiple categories, and the formula is as follows: .

[0011] As a further optimization of the present invention, in step S21, the backbone network module adopts the backbone network structure of the YOLOv5 model, and replaces the original cross-stage partial connection module with a context feature fusion module. The design of the context feature fusion module refers to the idea of ​​the cross-stage partial connection module and the efficient layer aggregation network in YOLOv7, so that while ensuring lightweight, richer gradient flow information is obtained; the neck network module adopts the neck network structure of the YOLOv5 model, and replaces the original cross-stage partial connection module with a context feature fusion module, and deletes the convolution module before performing the upsampling operation in the path aggregation-feature pyramid network module, and directly inputs the feature maps of the backbone network module at different stages into the upsampling.

[0012] As a further optimization of the present invention, the V5 loss function is composed of three loss functions: classification loss (cls_loss), regression loss (box_loss) and confidence loss (obj_loss); and the improved YOLOv8 model only uses the Varifocal Loss classification loss function. The main improvement is that an asymmetric weighted operation is proposed to solve the problem of imbalance between positive and negative samples. The calculation is shown as follows: , Among them, p refers to the predicted Iou-aware classification score, that is, the joint representation of the target in terms of confidence and positioning accuracy; q refers to the target score. For positive samples, q is the IOU value between the bounding box and the actual real box, and the BCE loss function is used. When it is a negative sample, q=0; the regression loss is to add DFL on the basis of Ciou Loss, that is, the regression loss function in the form of Ciou Loss+DFL, which uses the idea of ​​cross entropy function, which mainly enables the network to focus on the vicinity of the target position faster.

[0013] As a further preference of the present invention, in step S21, the BRA double-layer routing attention mechanism filters irrelevant key value pairs in the coarse area of ​​the feature map, that is, by constructing and pruning the region-level directed graph, and then applying fine-grained token-to-token attention in the union of the regions to achieve the filtering operation. After an image passes through the SPPF module of YOLOv8, the feature map at this moment is obtained by operating the original image 32 times. At this time, the feature map size is the smallest, so adding a convolution module and a BRA module at this position can capture the subtle information of the feature map, enhance the association of image context information, and improve the ability to extract small target features in the feature map.

[0014] As a further preferred embodiment of the present invention, step S22 loads the pre-trained weight yolov8s.pt and uses the collected labeled data to train the model; during the training process, the learning rate is set to 0.01, the batch size is 16, and the number of iterations is 100.

[0015] As a further preferred embodiment of the present invention, in step S23, the trained model is applied to the actual image or video stream for detection, and the method of comprehensive analysis of video signal + photoelectric signal is adopted. The calculation method of the detection result result is as follows: , in, , are the results of image target detection and photoelectric signal detection, respectively. , is the coefficient, is the classification threshold.

[0016] Compared with the prior art, the present invention has the following advantages or beneficial effects: The present invention provides a large vehicle right blind spot pedestrian collision prevention detection method based on an improved YOLOv8 model, with more accurate detection results and faster response time, avoiding the possibility of collision between large vehicles and pedestrians, reducing the occurrence of accidents, thereby reducing the cost of accident handling and compensation, and reducing the economic losses of enterprises. By testing and training the improved YOLOv8 model, the average accuracy of the detection results can be improved: the feature reorganization upsampling module in the model introduces the BRA double-layer routing attention mechanism and the CARAFE lightweight upsampling operator. The BRA mechanism effectively filters out irrelevant features by constructing and pruning a regional-level directed graph, so that the model can focus on small target feature information, improve the effectiveness of the feature graph, and reduce the complexity of subsequent calculations, so that the model reduces unnecessary calculation burden while ensuring performance, and then applies fine-grained token-to-token attention in the union of regions to implement filtering operations, which can capture more subtle feature relationships and better understand the contextual relationship between features. When an image passes through the SPPF module of YOLOv8, the feature map at this moment is obtained by downsampling the original image 32 times. At this time, the feature map size is the smallest. Therefore, adding a convolution module and a BRA module at this position can capture the subtle information of the feature map, enhance the association of image context information, and improve the ability to extract small target features in the feature map. CARAFE is a feature reorganization upsampling method based on content perception. Compared with linear interpolation upsampling, it processes the input feature map in blocks and uses convolution operations to reorganize the features of each block. This method can enhance the perception range, so that the upsampled feature map can integrate more context information and maintain the details and edge information of the original feature map, thereby improving the quality of upsampling and the ability to express features. As a lightweight upsampling operator, CARAFE reduces the computational and memory requirements of the model while retaining high performance, and is suitable for running in resource-limited environments. These advantages enable the BRA and CARAFE modules to significantly improve the performance of the YOLOv8 model in target detection tasks, especially small target detection and complex scene analysis.

[0017] The video set standard method (unified time, unified scene, unified route, unified alarm situation) is used to replace the manual method, which is more scientific, standardized, efficient and safe. The photoelectric combined with video method is used to judge the response time, replacing the manual timing, which is more scientific, standardized and accurate. When a pedestrian alarm occurs in a certain area of ​​the detected video, the alarm signal sent by the product is received, and the alarm response time is obtained by calculating the time difference between the two, meeting the judgment requirement of the alarm within 1s. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0019] Figure 1 This is a flowchart of pedestrian detection based on the improved YOLOv8 model; Figure 2 Schematic diagram of the overall structure of the improved YOLOv8 model; Figure 3 This is a structural diagram of the components of the backbone network module, feature reorganization upsampling module and Neck module. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0021] Please refer to the attached Figure 1 , this scheme provides a method for detecting pedestrians in the right blind spot of a large vehicle, including the following steps: S1. Data preprocessing: S11. Collect images of pedestrian warnings as a data set; pedestrian warnings include daytime positive pedestrian warnings, daytime reverse pedestrian warnings, nighttime positive pedestrian warnings, and nighttime reverse pedestrian warnings. 2,000 images are collected for each type of warning. The four types of warnings include static independent pedestrian scenes, umbrella pedestrian scenes, bicycle pedestrian scenes, raincoat bicycle pedestrian scenes, raincoat electric vehicle pedestrian scenes, and cargo electric vehicle pedestrian scenes. Each scene is designed with 15 alarm points. Since these alarm points need to distinguish between the front and back sides of pedestrians, it is necessary to shoot the round-trip pedestrian scene once. The initial plan estimates 100-200 alarm points. A waterproof camera is used for shooting, the video format is 1080P, and the shooting content is arranged completely according to the actual test scene. The personnel, clothing, venue, light intensity, etc. in various scenes are completely unified.

[0022] S12. Scale and crop the image to meet the input requirements of the improved YOLOv8 model, and then use random rotation, flipping, and brightness adjustment methods to increase the diversity of image data and enhance the robustness of the model.

[0023] S13, annotating the target area of ​​the image to obtain annotation data, and dividing the image into a training set and a test set; the ratio of the training set to the test set is 4:1, and 6400 images are randomly selected as the training set and 1600 images are randomly selected as the test set.

[0024] S2. Model training and application: S21. Use the test set to test the improved YOLOv8 model and save the model parameters with the highest average recognition accuracy, that is, the optimal model parameters.

[0025] like Figure 2 As shown in the figure, the improved YOLOv8 model includes: Backbone module, feature reorganization upsampling module and Neck module.

[0026] The backbone network module includes CBS module, C2f module, SPPF module, and the Neck module includes Connect module, C2f module, CARAFE module, and CBS module. The structure of each module is as follows: Figure 3 As shown. The CBS module includes Conv, BN, and SILU, which represent the convolution layer, batch normalization layer, and SiLU activation function, respectively, and are used to extract features and maintain the stability of information flow; the C2f module includes Split, Bottleneck, and Connect, which represent input channel division, deep feature extraction, and channel fusion, respectively. The C2f module is used for deep feature extraction and fusion in Backbone and Neck of YOLOv8; the SPPF module increases the receptive field while maintaining computational efficiency through multi-level pooling operations, which can help the model better understand the characteristics of the target at different scales; the CARAFE module redistributes feature values ​​by learning adaptively allocated weights, which is more advantageous than traditional upsampling methods (such as bilinear interpolation, deconvolution, etc.). It should be noted that the Backbone module adopts the backbone network structure of the YOLOv5 model, replacing the original C3 module with the C2f module. The design of the C2f module refers to the idea of ​​the C3 module and the ELAN in YOLOv7, so that while ensuring lightweight, it can obtain richer gradient flow information; the Neck module adopts the Neck structure of the YOLOv5 model, replacing the original C3 module with the C2f module, and deleting the convolution module before the upsampling operation in PAN-FPN, and directly inputting the feature maps of Backbone at different stages into the upsampling. The V5 loss function is composed of three loss functions: classification loss (cls_loss), regression loss (box_loss) and confidence loss (obj_loss); while the improved YOLOv8 model only uses the Varifocal Loss classification loss function. Its main improvement is the asymmetric weighted operation proposed to solve the problem of imbalance between positive and negative samples. The calculation is shown as follows: , Among them, p refers to the predicted Iou-aware classification score, that is, the joint representation of the target in terms of confidence and positioning accuracy; q refers to the target score. For positive samples, q is the IOU value between the bounding box and the actual real box, and the BCE loss function is used. When it is a negative sample, q=0; the regression loss is to add DFL on the basis of Ciou Loss, that is, the regression loss function in the form of Ciou Loss+DFL, which uses the idea of ​​cross entropy function, which mainly enables the network to focus on the vicinity of the target position faster.

[0027] The feature reorganization upsampling module includes a convolution-batch normalization-activation function module, a lightweight upsampling operator, and a two-layer routing attention module. The composition structure of each module is as follows: Figure 3 As shown in the figure, the Bi-level Routing Attention module first divides the feature map into S×S non-overlapping regions, and then uses the fully connected layer to project the input to obtain Q, K, and V, and calculates the mean values ​​of Q and K respectively. , , and then filtered and collected by the Gather module and , and finally to Q, and Apply fine-grained token-to-token attention to get the output of the two-layer routing attention module. The feature reorganization upsampling module introduces the BRA two-layer routing attention mechanism and the CARAFE lightweight upsampling operator. The image is sampled by the SPPF module of the YOLOv8 model to obtain the feature map. The BRA two-layer routing attention mechanism filters the irrelevant key value pairs in the coarse area of ​​the feature map, and then the CARAFE lightweight upsampling operator is used for block processing, and the convolution operation is used to reorganize the features of each block. The specific steps are: The improved YOLOv8 model is tested using the test set to calculate precision, recall, and average accuracy. Precision refers to the percentage of samples predicted as positive examples among all positive examples, and recall refers to the percentage of samples predicted as positive examples among all predicted examples. The calculation formulas for precision and recall are as follows: , Among them, TP (True Positive) refers to the number of samples correctly predicted by the model as positive (i.e., both the predicted result and the actual label are positive), FP (False Positive) refers to the number of samples incorrectly predicted by the model as positive (i.e., the predicted result is positive, but the actual label is negative), TN (True Negative) refers to the number of samples correctly predicted by the model as negative (i.e., both the predicted result and the actual label are negative), and FN (False Negative) refers to the number of samples incorrectly predicted by the model as negative (i.e., the predicted result is negative, but the actual label is positive).

[0028] With recall as the horizontal axis and precision as the vertical axis, a PR curve is formed as the threshold changes. The area enclosed by the PR curve and the horizontal and vertical axes is the average precision (AP), and the formula is as follows: The corresponding detection indicator mean average precision (mAP) is obtained by averaging multiple categories, and the formula is as follows: S22. Load the optimal model parameters into the improved YOLOv8 model, select YOLOv8 and improved YOLOv8 to detect pedestrians, load the official pre-trained weights yolov8s.pt, and train the model using the collected labeled data. During the training process, the learning rate is set to 0.01, the batch size is 16, and the number of iterations is 100. The training results are shown in Table 1.

[0029] Table 1 Offline training results of target detection model

[0030] S23. Apply the trained model to the actual image for detection, and record the response time and detection results.

[0031] The method of comprehensive judgment of video signal + photoelectric signal is adopted, and the calculation method of detection result result is as follows: , in, , are the results of image target detection and photoelectric signal detection, respectively. , is the coefficient, is the classification threshold.

[0032] The response time and detection results were recorded, and it was calculated that the average response time of this method was 513ms, and the average detection accuracy reached 98.6%.

[0033] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A large vehicle right blind spot pedestrian anti-collision detection method based on an improved YOLOv8 model, characterized in that: The steps include: S1. Data preprocessing: S11. Collect images of pedestrian warning situations as a data set. Pedestrian warning situations include daytime positive pedestrian warning situations, daytime reverse pedestrian warning situations, nighttime positive pedestrian warning situations, and nighttime reverse pedestrian warning situations. Each type of warning situation collects n images, where n≥1600. S12, standardize and enhance the image data; S13, annotating the target area of ​​the image to obtain annotation data, and dividing the image into a training set and a test set; the ratio of the training set to the test set is 4:1; S2. Model training and application: S21. Use the test set to test the improved YOLOv8 model, and save the model parameters when the average recognition accuracy is the highest, that is, the optimal model parameters; The improved YOLOv8 model includes: backbone network module, feature reorganization upsampling module and Neck module; The feature reorganization upsampling module introduces a double-layer routing attention mechanism and a lightweight upsampling operator. The spatial pyramid pooling module of the YOLOv8 model is used to sample the image to obtain a feature map. The double-layer routing attention mechanism filters irrelevant key value pairs in the coarse area of ​​the feature map, and then divides the image into blocks through a lightweight upsampling operator. The convolution operation is used to reorganize the features of each block. S22, load the optimal model parameters into the improved YOLOv8 model, and train the model using the labeled data; S23. Apply the trained model to the actual image for detection, and record the response time and detection results.

2. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1 is characterized in that: In step S11, the four types of police situations all include a static independent pedestrian scene, an umbrella pedestrian scene, a bicycle pedestrian scene, a bicycle pedestrian wearing a raincoat scene, an electric vehicle pedestrian wearing a raincoat scene, and a cargo electric vehicle pedestrian scene.

3. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1 is characterized in that: In step S12, the image is scaled and cropped to meet the input requirements of the improved YOLOv8 model, and then the random rotation, flipping, and brightness adjustment methods are used to increase the diversity of image data and enhance the robustness of the model.

4. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1 is characterized in that: The specific steps of step S21 are: The improved YOLOv8 model is tested using the test set to calculate precision, recall, and average accuracy. Precision refers to the percentage of samples predicted as positive examples among all positive examples, and recall refers to the percentage of samples predicted as positive examples among all predicted examples. The calculation formulas for precision and recall are as follows: ， Among them, TP refers to the number of samples correctly predicted by the model as positive, FP refers to the number of samples incorrectly predicted by the model as positive, TN refers to the number of samples correctly predicted by the model as negative, and FN refers to the number of samples incorrectly predicted by the model as negative. With recall as the horizontal axis and precision as the vertical axis, a PR curve is formed as the threshold changes. The area enclosed by the PR curve and the horizontal and vertical axes is the average accuracy. The formula is as follows: ， The corresponding average accuracy of the detection index is obtained by averaging multiple categories, and the formula is as follows: 。 5. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1 is characterized in that: In step S21, the backbone network module adopts the backbone network structure of the YOLOv5 model, replaces the original cross-stage partial connection module with a context feature fusion module, and the design of the context feature fusion module adopts the cross-stage partial connection module and combines the idea of ​​the efficient layer aggregation network in YOLOv7; the neck network module adopts the neck network structure of the YOLOv5 model, replaces the original cross-stage partial connection module with a context feature fusion module, and deletes the convolution module before performing the upsampling operation in the path aggregation-feature pyramid network module, and directly inputs the feature maps of the backbone network module at different stages into the upsampling.

6. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 4 is characterized in that: In step S21, the loss function used in the improved YOLOv8 model is replaced with the Varifocal Loss classification loss function. The problem of imbalance between positive and negative samples is solved through asymmetric weighted operations. The calculation is shown as follows: ， Among them, p refers to the predicted Iou-aware classification score, that is, the joint representation of the target in terms of confidence and positioning accuracy; q refers to the target score. For positive samples, q is the IOU value between the bounding box and the actual real box, and the BCE loss function is used. When it is a negative sample, q=0; the regression loss is to add DFL on the basis of Ciou Loss, that is, the regression loss function in the form of Ciou Loss+DFL.

7. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1 is characterized in that: In step S21, the two-layer routing attention mechanism filters irrelevant key-value pairs in the coarse area of ​​the feature map, that is, by constructing and pruning a region-level directed graph, and then applying fine-grained token-to-token attention in the union of regions to achieve the filtering operation.

8. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1 is characterized in that: Step S22 loads the pre-trained weights yolov8s.pt and uses the collected labeled data to train the model; during the training process, the learning rate is set to 0.01, batch size is 16, and number of iterations is 100.

9. The large vehicle right blind spot pedestrian anti-collision detection method based on the improved YOLOv8 model according to claim 1, characterized in that: In step S23, the trained model is applied to the actual image or video stream for detection, and the method of comprehensive analysis of video signal + photoelectric signal is adopted. The calculation method of the detection result result is as follows: in, , are the results of image target detection and photoelectric signal detection, respectively. , is the coefficient, is the classification threshold.

Images