Urban space mixed traffic participant detection model based on improved YOLOv8n algorithm

Abstract

The invention discloses an urban space mixed traffic participant detection model based on an improved YOLOv8n algorithm, and relates to the technical field of intelligent traffic detection. In an input stage, an input picture data set is enhanced; using a space-to-depth non-cross-row convolution module to replace an original convolution layer of the convolutional neural network in the feature extraction stage, the feature fusion stage and the target detection stage, and performing down-sampling on the feature map; a lightweight coordinate attention mechanism module is added in the feature fusion stage, position information is embedded into channel attention, and cross-channel information and direction position sensing information are captured from a feature map; and in the target detection stage, an EIoU loss function is adopted to measure the difference between a prediction result and a real result. According to the method, the feature extraction capability of low-resolution small targets of traffic participants is improved, the accuracy, real-time performance and stability of detection of multiple types of traffic participants are improved, and reliable theoretical support is provided for intelligent safety detection and risk identification.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent traffic detection, and more specifically, to an urban space mixed traffic participant detection model based on an improved YOLOv8n algorithm. Background Art

[0002] With the continuous acceleration of the urbanization process and the diversification of transportation tools, the types and quantities of traffic participants (pedestrians, motor vehicles, non-motor vehicles, etc.) have shown exponential growth and are often in a complex spatial mixed state, which makes the real urban traffic environment become increasingly complex and the traffic safety problem more severe. The emergence of intelligent object detection technology has alleviated the traffic safety problem. In the spatial mixed state, being able to accurately identify and monitor multi-category traffic participants quickly and accurately in complex scenarios such as dense, occluded, low-resolution small targets, and bad weather is of great significance for the safe driving of intelligent connected vehicles and the real-time safety control of urban intelligent traffic monitoring systems.

[0003] In recent years, because deep learning can learn effective features from data, it has gradually replaced traditional machine learning algorithms and become the mainstream traffic participant detection algorithm. Single-stage detection algorithms such as the You Only Look Once (YOLO) series, RetinaNet, and Single Shot MultiBox Detector (SSD), etc., are more superior to two-stage algorithms in terms of detection real-time performance and deployment convenience due to the one-step implementation from the input image to obtaining the bounding box position and category, and are widely used for the detection of traffic and pedestrians. Xia et al. proposed a moving target detection model based on an improved YOLOv3 algorithm, and through the improvement of the algorithm backbone network, achieved effective real-time detection of vehicles and personnel. Shu et al. built a small target pedestrian detection model based on an improved YOLOv4 algorithm, and this model obtained good robustness and feature discriminability. Wang et al. achieved an identification accuracy of about 93% for invasive target foreign objects along the traffic by strengthening the fusion of the backbone and neck network layers of the YOLOv5s algorithm. Zhou et al. proposed a traffic vehicle monitoring model based on an improved YOLOv6 algorithm for the problems of difficult monitoring and high false detection rate of traffic vehicles in harsh environments, and greatly improved the detection performance of the model. Wang et al. proposed a dense pedestrian detection model based on an improved YOLOv8 algorithm, and by introducing an attention mechanism strategy, the feature extraction ability and detection accuracy of the model were greatly improved. YOLOv8 is the latest detection model in the YOLO series, which has a significant improvement in detection speed and accuracy. At the same time, due to its convenience for lightweight device deployment, it is very suitable for the detection of traffic participants in the spatial mixed scenario after improvement.

[0004] The above research has achieved good results in the detection of different categories of traffic targets, but there is less research on the detection of multi-category traffic participants, especially the precise and real-time detection research in the complex background of mixed traffic in urban spaces is relatively rare. Summary of the Invention

[0005] The present invention aims to provide a detection model for traffic participants in mixed urban spaces based on an improved YOLOv8n algorithm, which can solve the above problems.

[0006] To solve the above problems, the technical solutions adopted by the present invention are as follows:

[0007] The present invention provides a detection model for traffic participants in mixed urban spaces based on an improved YOLOv8n algorithm, including an input stage, a feature extraction stage, a feature fusion stage, a target detection stage, and an output stage of the YOLOv8n model. In the input stage, the input picture dataset is enhanced; a space-to-depth non-stride convolution module is used to replace the original convolution layers of the convolutional neural network in the feature extraction stage, feature fusion stage, and target detection stage to downsample the feature map; a lightweight coordinate attention mechanism module is added in the feature fusion stage to embed position information into channel attention to capture cross-channel information and direction position perception information from the feature map; in the target detection stage, the EIoU loss function is used to measure the gap between the prediction result and the real result.

[0008] As a further description of the above technical solution, in the input stage, for a certain input picture data, one of the four data enhancement transformation methods of random geometric transformation, contrast adjustment, resolution reduction, and noise addition is used to enhance it.

[0009] As a further description of the above technical solution, when the space-to-depth non-stride convolution module downsamples the feature map, let any intermediate feature map with a size of C1×S×S in length, width, and height be X, then the sub-feature map sequence slice is:

[0010]

[0011] In the formula, scale is the downsampling ratio factor used to determine the size of the sub-feature map; f(x,y) represents the sub-feature map sequence slice, where x and y are the starting positions of the slice.

[0012] As a further description of the above technical solution, the working method of the lightweight coordinate attention mechanism module includes: the feature map input into the module first performs one-dimensional global pooling operations through average pooling in the X direction and average pooling in the Y direction, aggregating the input features into two independent direction-aware feature maps respectively; then encodes the perception feature maps into two attention maps through batch normalization and non-linear activation functions; finally, through separation and direction weighting, each attention map captures the long-range dependence of the input feature map along a spatial direction.

[0013] As a further description of the above technical solution, the formula of the EIoU loss function is defined as follows:

[0014] L EIoU = L IoU + L dis + L asp

[0015] L IoU = 1 - IoU

[0016]

[0017] Among them, IoU is the intersection over union, L EIoU is the EIoU loss, L IoU is the IoU loss, L dis is the center distance loss, L asp is the width and height loss, w c and h c are the width and height of the smallest bounding box covering the two boxes respectively.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1) By integrating the SPD-Conv module and the CA module into the YOLOv8n model, the detection ability and feature extraction ability of the model for low-resolution small targets are significantly improved. The improved data input enhancement strategy combined with the bounding box EIoU loss function effectively improves the robustness and real-time performance of traffic participant detection.

[0020] 2) Experimental results prove that the SEC-YOLO algorithm has high accuracy and real-time performance in detecting urban multi-class traffic participants, and its detection accuracy is better than algorithms such as YOLOv8n, YOLOv5n, and Faster R-CNN. In terms of dataset detection, the recognition speed reaches 93.8 frames / s, and the mean average precision reaches 94.5%; in terms of real-scene detection, the accuracy rate reaches 95%.

[0021] To make the above objects, features, and advantages of the present invention more obvious and understandable, specific embodiments of the present invention are hereinafter given, and detailed descriptions are made in conjunction with the accompanying drawings as follows. Brief Description of the Drawings

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as a limitation of the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative efforts.

[0023] Figure 1 is the structural framework diagram of the urban space mixed traffic participant detection model (SEC-YOLO);

[0024] Figure 2 is an example of SPD-Conv downsampling when the magnification scale = 2;

[0025] Figure 3 is the structural schematic diagram of the CA module;

[0026] Figure 4 is the bounded box regression diagram of the CIoU loss function;

[0027] Figure 5 is a partial image of the collected dataset;

[0028] Figure 6 is the comparison of the effects of partial randomly introduced data augmentation;

[0029] Figure 7 is the performance comparison between the SEC-YOLO model and the YOLOv8n model described in the present invention;

[0030] Figure 8 is the comparison of the mAP@0.5 / % and R results of each model;

[0031] Figure 9 is the comparison diagram of the detection results of some models;

[0032] Figure 10 is the real-time detection result diagram of some autonomous driving experiments;

[0033] Figure 11 is the comparison diagram of the detected number of people and the actual number of people. Detailed Embodiments

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0035] The detection process of the YOLOv8n model includes an input stage, a feature extraction (Backbone) stage, a feature fusion (Neck) stage, an object detection (Head) stage, and an output stage. In the feature extraction stage, traffic participant target features are extracted from the input images; in the feature fusion stage, the extracted target features are fused and the target features detected in each layer are combined; in the object detection stage, the detection results are output and some are screened to remove clutter and redundancy.

[0036] The present invention provides a detection model for mixed traffic participants in urban space, which has made four key improvements based on the above YOLOv8n model. The detailed structure is as Figure 1 shown. First, in the image input stage, geometric transformation and pixel transformation strategies are used to enhance the input image dataset to prevent overfitting and improve robustness and generalization; second, the original all-convolution structures (i.e., the original convolutional layers in the feature extraction stage, feature fusion stage, and object detection stage) are replaced with a Space-to-Depth Convolution (SPD-Conv) module to improve the feature extraction ability for small objects with low resolution; subsequently, a lightweight coordinate attention mechanism (Coordinate attention, CA) is introduced into the neck fusion network structure to improve the feature extraction ability in the Neck network structure; finally, the Efficient Intersection over Union (Efficien IoU, EIoU) is used to replace the original loss function to optimize the calculation of the bounding box. The finally obtained detection model for mixed traffic participants in urban space is named SEC-YOLO.

[0037] The SPD-Conv module, CA module, and EIoU loss function of the SEC-YOLO model are described below.

[0038] 1. SPD-Conv module

[0039] When traffic monitoring systems and intelligent connected vehicles detect multiple types of urban traffic participants, there are often situations where the detection targets are small objects or have low resolution. At this time, the Convolution Neural Networks (CNN) of YOLOv8n will suffer from the loss of fine-grained information and poor feature learning, and the detection accuracy will be greatly reduced. To address this problem, the present invention introduces the SPD-Conv module to replace each strided convolution layer and pooling layer of the CNN. The SPD-Conv module can downsample the feature map without losing learnable information, greatly compensating for the defect of the original model in the feature detection performance of traffic participants with low resolution and small targets. The SPD component in the SPD-Conv module downsamples the entire feature map of the CNN through image conversion technology, and the principle is as follows.

[0040] Let the intermediate feature map of any size S×S×C1 be X, then the slice of the sub-feature map sequence is as follows:

[0041]

[0042] In the formula, scale is the downsampling ratio factor, which is used to determine the size of the sub-feature map; f(x, y) represents the slice of the sub-feature map sequence, where x and y are the starting positions of the slice. X is any given feature map, and the sub-map f(x, y) is composed of all feature maps to form features Figure X (i, j), where i and j represent features Figure X The index positions on it identify the position of a sub-region or an element of a feature map in the feature map. Each sub-map downsamples X by a scale factor. This mapping enables the convolutional layer to operate at different dilation rates, thereby obtaining a wider field of view at different scales. Figure 2 As an example of SPD-Conv downsampling when scale = 2, four sub-maps can be obtained, and their shapes are (S / 2, S / 2, C1). Subsequently, X is downsampled by a factor of 2, and the final obtained channel dimension will increase by a scale factor of 2.

[0043] By referring to the SPD-Conv module, YOLOv8n can obtain a wider receptive field at different scales. In the mixed traffic state in urban spaces, it greatly enhances the model's feature extraction ability for small targets and low-resolution targets of multi-class traffic participants, thereby improving the detection accuracy.

[0044] 2. CA Module

[0045] In the urban intersections with mixed traffic, the collected images often contain multiple types of traffic participants and are dense in quantity. The traditional YOLO series will accumulate a large amount of redundant invalid information due to the convolutional iteration of multi-target information, thus drowning out the target information. To address the above problems, the present invention introduces a lightweight coordinate attention mechanism CA module. By embedding position information into channel attention, this module not only captures cross-channel information but also includes direction-aware and position-aware information, enabling the urban space mixed traffic participant detection model to more accurately locate and identify the target area, while also avoiding a large amount of computational overhead. Its implementation principle is as follows Figure 3As shown in the figure, where C represents the number of channels, W represents the width, H represents the height, and 1 represents the dimension. The input image first undergoes a one-dimensional global pooling operation through X Avg Pool (average pooling in the X direction) and Y Avg Pool (average pooling in the Y direction), aggregating the input features into two independent direction-aware feature maps respectively; subsequently, the perceived feature maps are encoded into two attention maps through BatchNorm (batch normalization) and Non-linear (non-linear activation function); finally, through split separation and direction weighting, each attention map successfully captures the long-range dependencies of the input feature map along a spatial direction.

[0046] Position perception in the lightweight coordinate attention mechanism (CA) enhances the model's understanding of the specific position of the target in space by embedding position information into the channel attention mechanism. In traditional convolutional neural networks, when the model extracts information through feature maps, it usually does not consider the specific position of these features in the image. The CA module solves this problem by extracting direction-aware features and combining them with the channel attention mechanism, enabling the model to not only recognize the target but also perceive the spatial position and relative relationship of the target. Specifically, the CA module extracts position features in the X direction and Y direction respectively through a one-dimensional global pooling operation. These features represent the spatial distribution of the target in the image, and the information is fused with other features through splicing and convolution, thereby enhancing the model's ability to perceive the target position. In this way, the model can weight the features according to the position of the target in the image, making the detection more accurate, especially in complex urban traffic environments, where it can better handle the positioning and recognition of targets. At the same time, since the CA module adopts a lightweight design, it avoids unnecessary computational overhead and can improve the running efficiency of the model while ensuring accuracy.

[0047] The incorporation of the CA module can not only effectively improve the recognition accuracy of the key information of traffic participants by the model, but also, due to its lightweight structural characteristics, hardly adds any computational overhead to the device during operation.

[0048] 3. EIoU Loss Function

[0049] The loss function is used to measure the gap between the prediction result and the true result. The traditional YOLOv8 series all use the Complete IoU (CIoU) loss function as the calculation of the boundary loss. However, as Figure 4 shown, the aspect ratio of the CIoU loss function is a relative value, which is prone to imbalance between irregular target frames, resulting in a decline in the detection stability of the model. To address this problem, the present invention uses the EIoU loss function to replace the CIoU loss function in the YOLOv8n model.

[0050] The EIoU loss function replaces the aspect ratio by calculating the difference values of width and height respectively, and solves the problem of target frame imbalance. In the present invention, the formula of the EIoU loss function is defined as follows:

[0051] L EIoU = L IoU + L dis + L asp (2)

[0052] L IoU = 1 - IoU (3)

[0053]

[0054] Among them, the EIOU loss function includes three parts: IoU loss, center distance loss L dis and width-height loss L asp , w c and h c are the width and height of the smallest bounding box covering two boxes respectively, ρ represents the center distance, b gt and b pre represent the center positions of the ground truth box and the predicted box respectively, w gt and h gt are the width and height of the ground truth box respectively, IoU is the intersection over union of the predicted box and the ground truth box, and w and h are the width and height of the predicted box respectively.

[0055] The EIoU loss function effectively avoids the failure caused by different aspect ratios when detecting traffic participants. Therefore, the application of the EIOU loss function can better ensure the stability and real-time performance of intelligent connected vehicles and traffic monitoring systems when detecting traffic participants.

[0056] The SEC-YOLO model described in the present invention is trained based on the BNS dataset. The establishment and preparation process of the BNS dataset are as follows:

[0057] The dataset used in the present invention comes from two sources: the publicly available dataset on the platform and the field-collected dataset, to ensure the input of rich driving scenarios and improve the generalization ability and robustness of the model. Since traffic participants need to be in the traffic state of spatial mixing, the parts that meet the scenarios in the BDD100K dataset and the NEXET dataset are selected, and the self-built dataset is collected and labeled to supplement the missing scenarios of the dataset. Finally, the three datasets are fused into the BNS (BDD100K, NEXET, Self-built) dataset.

[0058] The self-built dataset was collected from multiple streets and intersections in Chongqing, China. The acquisition device was the Canon r6 Mark II, and a total of 42 videos with a duration of 10 minutes were collected. The video types include videos of 13 different scenarios such as intersections with dense mixed traffic of people and vehicles, zebra crossings on urban traffic roads, non-motorized lanes in rainy days, and streets with mixed traffic of people and vehicles at night. Subsequently, key frames were extracted through OpenCV, and some of the extracted images are as shown in Figure 5 as shown. Finally, the LabelImg tool was used to label the types and positions of traffic participants in the YOLO format.

[0059] To improve the recognition performance of the improved model for the characteristics of traffic participants, the present invention adopts geometric transformation and pixel transformation enhancement strategies. At the data input stage, four data enhancement transformation methods, namely geometric transformation, contrast adjustment, resolution reduction, and noise addition, are randomly introduced into the dataset, as shown in Figure 6 shown. In the figure, the left attached figures are all the figures before enhancement, and the right attached figures are the figures after enhancement. (a) is the enhancement method of rotation angle, (b) is the enhancement method of adjusting contrast, (c) is the enhancement method of reducing the resolution of the target to be detected, and (d) is the enhancement method of adding noise. The above methods enrich the input methods of data training and can effectively improve the robustness and generalization ability of the model for traffic participant detection. After the useful scene screening and random data enhancement of the BNS dataset, a total of 216,512 labeled images were obtained. To ensure the randomness of dataset segmentation and reduce the risks of overfitting and noise influence, the 5-fold cross-validation method is adopted for the BNS dataset. The method is to randomly divide the dataset into 5 parts, take 4 parts as the training set each time, and the remaining one part as the validation set, and the cross-validation can be repeated 5 times.

[0060] In the specific implementation of the present invention, the experimental device used is a computer, and its configuration is shown in Table 1. All the experiments involved are carried out in this environment. During the training process, the number of training rounds is set to 300, the initial learning rate is 0.002, the weight decay is 0.0005, and the batch size is set to 64.

[0061] Table 1 Experimental environment configuration

[0062]

[0063] The floating-point operation count (FLOPs) that measures the computational complexity of the model and the frames per second (FPS) that measures the real-time performance of the model are selected as the evaluation indicators of the model performance. Precision (P), Recall (R), and Mean Average Precision (mAP) are used as the measurement indicators of the model accuracy. The corresponding calculation processes are as follows:

[0064]

[0065] In the formula, Q n is the total number of detection views; T is the total time used; TP is the number correctly recognized; FP is the number misrecognized; FN is the number not recognized or misrecognized.

[0066] The mean average precision and the boundary loss of the SEC-YOLO model obtained in the present invention were compared with the original model. The number of training rounds was 0 - 300 rounds, and the results are as Figure 7 shown. As can be seen from Figure (a), after about 60 rounds of iteration, the recognition accuracy of the SEC-YOLO model for traffic participants has been greatly improved; as can be seen from Figure (b), after about 50 rounds of iteration, the boundary loss curve has tended to be stable, and the SEC-YOLO model converges faster and more stably. Based on the above analysis, the SEC-YOLO model obtained in the present invention performs better in the traffic participant detection task, which proves the feasibility of the improved model.

[0067] To verify the effectiveness of each module for the improved model, we designed an ablation experiment, and the results are shown in Table 2.

[0068] Table 2 Results of the ablation experiment

[0069]

[0070] As can be seen from the results in Table 2, when only the SPD-Conv module is used, mAP@0.5 / % is increased by 2.3%, indicating that the SPD-Conv module can improve the overall performance of the network model; when only the CA module is used, mAP@0.5 / % is increased by 1.9%, while there is no improvement in FLOPs, indicating that the CA module can improve the detection accuracy of the model with almost no computational overhead; when only the EIOU loss function is used, mAP@0.5 / % still has a slight increase, indicating that the EIOU loss function can improve the detection accuracy of the model; when the three model improvements are used simultaneously, mAP@0.5 / % and mAP@0.5 - 0.95 / % are increased by 4.7% and 11.9% respectively. It can be seen that the SEC-YOLO model described in the present invention has a significant improvement in the traffic participant detection performance compared with the original model.

[0071] To verify the performance of the SEC-YOLO model, we used six other mainstream object detection models in the comparative experiment of traffic participant object detection. The comparative experiment results are shown in Table 3 and Figure 8 as shown. As can be seen from Table 3, the number of frames transmitted by the SEC-YOLO model reaches 93.8 frames / s, and the detection real-time performance is the strongest; by Figure 9It can be seen that the SEC-YOLO model has the highest mAP@0.5 / % and R value, and the fastest convergence speed.

[0072] Table 3 Experimental results of comprehensive indicators of improved models and mainstream models

[0073]

[0074] We selected the baseline model YOLOv8n and the model SEC-YOLO described in the present invention for a traffic participant detection experiment comparison, and the results are as Figure 9 shown. Figure 9 In the figure, the text and numbers in the prediction box above the traffic participants represent the category and confidence level respectively, and the upper right corner of the picture is a partial enlarged view of the object detection. The results show that the baseline model YOLOv8n has certain missed detections and false detections in the detection of complex scenarios such as small targets in the distance, multi-category dense aggregation, and low-light environments, as can be seen from the partial enlarged view in the second column; the SEC-YOLO model can effectively detect in the above three environments, and the false detection rate is lower, as can be seen from the third column of the figure. In summary, the SEC-YOLO model can accurately identify some occluded or low-resolution targets in the spatial mixed traffic scenario, and has more advantages in detecting traffic participants in complex traffic scenarios.

[0075] We used the autonomous driving experimental test platform of Chengdu Panfeng Technology Co., Ltd. to import the SEC-YOLO model into the platform to detect traffic participants. The detection address was a certain street in Yibin City, China, and the detection results are as Figure 10 shown. In order to accurately measure the missed detection rate, the platform counted the number of people detected in real time during the detection, and the actual number was counted manually. The comparison of the obtained results is as Figure 11 shown. The results show that when at the mileage point of the intersection where the number of traffic participants surges, there will be a certain deviation between the detected number and the actual number, but the average accurate recognition rate can still reach more than 95%. In summary, it can be seen that the improved algorithm model of the present invention has good real-time detectability and low false detection rate for traffic participants on the actual street.

[0076] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A mixed traffic participant detection model for urban space based on an improved YOLOv8n algorithm, comprising an input stage, a feature extraction stage, a feature fusion stage, a target detection stage and an output stage of a YOLOv8n model, characterized in that: In the input stage, the input image dataset is enhanced; a space-to-depth non-cross-row convolution module is used to replace the original convolution layer of the convolutional neural network in the feature extraction stage, the feature fusion stage and the target detection stage to downsample the feature map; a lightweight coordinate attention mechanism module is added to the feature fusion stage to embed the position information into the channel attention, and the cross-channel information and the direction position perception information are captured from the feature map; in the target detection stage, the EIoU loss function is used to measure the gap between the predicted result and the actual result.

2. The urban space mixed traffic participant detection model according to claim 1 is characterized in that: In the input stage, for a certain input image data, one of four data enhancement transformation methods including geometric transformation, contrast adjustment, resolution reduction and noise addition is randomly used to enhance it.

3. The urban space mixed traffic participant detection model according to claim 1 is characterized in that: When the space-to-depth non-cross-row convolution module downsamples the feature map, let the intermediate feature map of arbitrary length, width and height size C1×S×S be X, then the sub-feature map sequence slice is: Where scale is the downsampling scale factor, which is used to determine the size of the sub-feature map; f(x, y) represents the sub-feature map sequence slice, where x and y are the starting positions of the slice.

4. The urban space mixed traffic participant detection model according to claim 1 is characterized in that: The working method of the lightweight coordinate attention mechanism module includes: the feature map input to the module is first subjected to one-dimensional global pooling operations through X-direction average pooling and Y-direction average pooling, respectively aggregating the input features into two independent direction-aware feature maps; then the perceptual feature map is encoded into two attention maps through batch normalization and nonlinear activation function; finally, through separation and direction weighting, each attention map captures the long-range dependency of the input feature map along a spatial direction.

5. The urban space mixed traffic participant detection model according to claim 1 is characterized in that: The formula of the EIoU loss function is defined as follows: L EIoU =L IoU +L dis +L asp L IoU =1-IoU Among them, L EIoU is the EIoU loss, L IoU is the IoU loss, L dis is the center distance loss, L asp is the width-height loss, w c and h c are the width and height of the smallest bounding box covering the two boxes, ρ represents the center distance, b gt and b pre Represent the center positions of the real box and the predicted box respectively, w gt and h gt are the width and height of the real box, IoU is the intersection over union of the predicted box and the real box, w and h are the width and height of the predicted box, respectively.

6. The urban space mixed traffic participant detection model according to claim 1 is characterized in that: The detection model is trained based on the BNS dataset, which is composed of image data from the platform's public dataset and field-collected datasets. The platform's public datasets include the BDD100K dataset and the NEXET dataset.

7. The urban space mixed traffic participant detection model according to claim 6 is characterized in that: The BNS dataset is randomly divided into 5 parts, 4 parts are taken each time as training sets, and the remaining part is used as validation set for cross-validation.

Images