Target detection method for shielded vehicles in urban road scene

Abstract

The invention discloses a target detection method suitable for a shielded vehicle in an urban road scene, and aims to solve the problem of confusion of a target and a background caused by various shielding in the fields of unmanned driving, intelligent transportation and the like so as to improve the target detection precision. According to the method, on the basis of a YOLOv8 network, a mixed attention mechanism is introduced into a backbone network, so that the channel and space information capturing capability in feature extraction is enhanced; meanwhile, a multi-scale feature adaptive fusion module is added to the neck network, features of different scales are fused through cross-layer information, and the detection capability of the network is improved; and an up-sampling module of YOLOv8 is improved, a dynamic hierarchical up-sampling strategy is adopted, dynamic hierarchical sampling is performed on input features, and the spatial resolution and semantic expression ability of the features are optimized. According to the method, the target detection precision is effectively improved, the interference of shelters is reduced, the high precision is ensured, the rapid detection is realized, and the method is particularly suitable for carrying out the target detection task of an automatic driving system in an urban road environment.

Description

Technical Field

[0001] The present invention belongs to the field of target detection, and mainly relates to a target detection method for occluded vehicles in urban road scenes. Background Art

[0002] With the rapid development of intelligent transportation and autonomous driving technologies, vision-based object detection technology is increasingly being used in urban road scenarios. Vehicle detection, as a core task in intelligent transportation systems, is not only used for real-time traffic flow monitoring, but also provides a perception basis for autonomous vehicles. However, there are a large number of occlusions in complex urban road scenarios, such as the front vehicle occluding the rear vehicle, and partial occlusion by buildings or road facilities, which puts higher requirements on the object detection algorithm.

[0003] Traditional vehicle detection methods are mostly based on feature engineering and rule design, such as background difference method, Haar feature detector or HOG+SVM model. These methods can achieve certain detection results under ideal conditions, but the detection accuracy often drops significantly when faced with realistic challenges such as illumination changes, occluded targets, and complex backgrounds. In addition, traditional methods lack a deep understanding of local features and are prone to false detection or missed detection, which limits the scope and effect of practical applications.

[0004] In recent years, the rapid development of deep learning technology has provided powerful tools for target detection tasks. Target detection algorithms based on convolutional neural networks (CNNs), such as Faster R-CNN, YOLO, and SSD, can learn rich image features through end-to-end training, significantly improving the accuracy of detection. However, in urban road occlusion scenarios, existing algorithms still face important challenges: on the one hand, occlusion will lead to incomplete target features, making it difficult for the model to accurately identify the target; on the other hand, the detection algorithm is more sensitive to background interference information in the occluded area, which may cause false detection or false alarms.

[0005] In view of the above problems, the present invention proposes a target detection method for occluded vehicles in urban road scenes. The method can effectively improve the target detection accuracy of the model for occluded vehicles in urban road scenes. Summary of the invention

[0006] In view of this, the technical problem to be solved by the present invention is to provide an improved occluded vehicle target detection method based on YOLOv8, which can improve the target detection accuracy of the YOLOv8 model for occluded vehicles in urban road scenes.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for detecting a target blocked by a vehicle in an urban road scene, the method specifically comprising the following steps:

[0009] Step 1) construct a dataset for occluded vehicle target detection in urban road scenes and preprocess the images in the dataset;

[0010] Step 2) inputting the preprocessed image to be detected into the improved vehicle detection model, extracting features of the image through the improved backbone network, and generating a feature map of the image to be detected to capture multi-level feature information of the vehicle;

[0011] Step 3) Using the extracted feature map as the input of the feature fusion module to fuse the feature map information of different scales;

[0012] Step 4) The fused high-quality feature map is input into the detection head, which processes the feature map, locates the position of the target and determines the category, thereby completing the target detection task;

[0013] Further, the step 1) specifically includes the following steps:

[0014] Step 11) Obtain diverse image data in real scenes to construct a data set, which contains various complex situations, such as partial occlusion, complete occlusion, lighting changes, complex background, etc.

[0015] Step 12) Preprocess the images in the dataset, including image normalization and resizing, to improve the robustness and generalization ability of the model under occlusion.

[0016] Further, the step 2) specifically includes the following steps:

[0017] Step 21) The preprocessed data set is sent to the improved backbone feature extraction network YOLOv8 for feature extraction. The hybrid attention mechanism (HA) is introduced into the backbone network. The hybrid attention mechanism combines the CA attention mechanism and the GAM attention mechanism, so that the model can not only improve the sensitivity to local targets, but also improve the adaptability to complex scenes through global information. For road scenes in practical applications, this enhanced attention mechanism can help the model maintain a high accuracy in different scenarios and improve the detection performance of the model;

[0018] Step 22) Output multi-scale feature maps in the backbone network, which contain local detail information and global semantic information of the vehicle as input for subsequent modules.

[0019] Further, the step 3) specifically includes the following steps:

[0020] Step 31) Replace the C2f module in the neck network with a Multi Scale Feature Adaptive Fusion Module (MSFAFM), so that the model can have better multi-scale target detection capabilities under complex backgrounds and diverse vehicle sizes, and improve the expression and fusion capabilities of image features, capture more fine-grained information, and effectively distinguish targets from backgrounds;

[0021] Step 32) Replace the upsampling module in the neck network with a dynamic hierarchical upsampling module (DHU). DHU can dynamically adjust the sampling position and range according to the different levels of input features, and dynamically sample and reorganize the low-level features containing detail information and the high-level information containing global semantic information, thereby improving the accuracy of feature expression.

[0022] Further, the step 4) specifically includes the following steps:

[0023] Step 41) The fused feature map is input into the detection head, which predicts the bounding box coordinates and category of each target through regression and classification operations. In addition, the detection results are post-processed to remove redundant boxes through non-maximum suppression to ensure the accuracy and uniqueness of the output results. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0025] Figure 1 is a flow chart of the present invention;

[0026] Figure 2 It is the overall model framework diagram of the present invention;

[0027] Figure 3 This is the structural diagram of the hybrid attention mechanism proposed in the present invention;

[0028] Figure 4 This is a structural diagram of the multi-scale feature adaptive fusion module proposed in the present invention;

[0029] Figure 5 This is a structural diagram of the dynamic hierarchical upsampling module proposed in the present invention. DETAILED DESCRIPTION

[0030] In order to make the technical solutions, advantages and purposes of the present invention more clear, the present invention will be further described in detail below in conjunction with embodiments and drawings.

[0031] The present invention provides a method for detecting a target blocked by a vehicle in an urban road scene. The method flow chart is as follows: Figure 1 As shown in the figure, the overall model framework diagram proposed by this method is as follows Figure 2 As shown, the method comprises the following steps:

[0032] Step 1) construct an occluded vehicle target detection dataset in urban road scenes;

[0033] Step 2) Construct a hybrid attention mechanism (HA);

[0034] Step 3) construct a backbone feature extraction network;

[0035] Step 4) constructing a multi-scale feature adaptive fusion module (MSFAFM);

[0036] Step 5) construct a dynamic hierarchical upsampling module (DHU);

[0037] Step 6) Constructing a target detection network for occluded vehicles in urban road scenes;

[0038] Step 7) training the target detection network;

[0039] Step 8) Test the object detection network.

[0040] Further, step 1) specifically includes the following steps:

[0041] Step 11) Use an open source website to obtain open source urban road scene images. The image dataset contains various types of vehicles on urban roads. Use LabelImg image annotation software to annotate each image with its target category and its real position, and generate an annotation file, which is used as a dataset together with the original image.

[0042] Step 12) The obtained data set is divided into a training set, a test set, and a validation set in a ratio of 8:1:1, which are used for training, validation, and testing of the target detection network, respectively.

[0043] Further, step 2) specifically includes the following steps:

[0044] Step 21) Figure 3As shown in the figure, the hybrid attention mechanism is composed of a global attention mechanism (GAM) and a coordinate attention mechanism (CA). Input feature F, where the channel attention module extracts channel information by performing MaxPool and AvgPool operations on the input features. These pooled features are passed to the shared multi-layer perceptron (MLP) to generate a channel attention map (M_c). Then, the channel attention map is multiplied element-wise with the input features to weight the input features and highlight important channel information. The channel refined features (F') are processed by the convolution layer to obtain spatial features. Through MaxPool and AvgPool operations, spatial information is extracted and a spatial attention map (M_s) is generated. The spatial attention map is multiplied element-wise with the channel refined features to highlight important areas in space. The CA module focuses on modeling the spatial dimension. It obtains feature information in both directions by changing the original average pooling to the maximum pooling operation in the height and width directions. Since the background on the road (such as road surface, trees, buildings, etc.) may be similar to the characteristics of the car in many cases, the use of maximum pooling can effectively suppress background interference. The concatenated result is then processed by the Conv2D convolution layer, and BatchNorm and nonlinear activation are performed to further refine the features. The hybrid attention mechanism dynamically adjusts the influence of GAM and CA through a learnable weight generation module DN-Weight, which extracts global statistical information of input features through global average pooling, and uses two layers of fully connected networks Fc1 and Fc2 to gradually reduce dimensions and map, and uses Softmax to normalize the weights to ensure that their sum is 1, thereby dynamically allocating weights to adapt to the distribution of input features. This weight controls the proportion of the two in the output. By combining the GAM and CA modules, the hybrid attention mechanism can more comprehensively capture the channel and spatial information of the features, thereby improving the performance of the model in complex tasks.

[0045] Further, step 3) specifically includes the following steps:

[0046] Step 31) Use the hybrid attention mechanism constructed in step 2 to combine the convolutional module consisting of the convolutional layer, batch normalization layer, and activation function layer, the CSP bottleneck structure composed of the CSP structure, and the maximum pooling module (SPPF) to build the backbone feature extraction network, such as Figure 2 As shown in Backbone.

[0047] Further, step 4) specifically includes the following steps:

[0048] Step 41) Figure 4As shown in the figure, the multi-scale feature adaptive fusion module consists of an adjustable kernel convolution (AKConv) with flexible adjustment of the convolution kernel size, an enhanced residual feature extraction module (ERFEM) through residual connection and deep convolution layer combination, and a bottleneck module combined with an attention mechanism. The variable kernel convolution includes an adaptive convolution kernel design, which dynamically adjusts the sampling position of the feature map through the learned offset to effectively capture multi-scale information; the enhanced residual feature extraction module selectively adds the input features after 1×1 convolution to the features processed by the variable kernel convolution. When RC=True, the residual connection is used, and when RC=False, the residual connection is not used, so that the network can more easily propagate gradients during training, avoiding the gradient vanishing problem, thereby accelerating the convergence speed, and avoiding the increase of the computational complexity of jump connections in unnecessary places, reducing the computational overhead.

[0049] In the multi-scale feature adaptive fusion module, the input feature map passes through a 1×1 convolution block and then is divided into two parts, M1 and M2, according to the channel dimension. M1 is output after passing through N ERFEM modules, and each ERFEM is connected in parallel with M2. The value of N depends on the model complexity. In the present invention, the value of N is equal to 2. Then the feature map output by each ERFEM is spliced ​​together with M2, and the spliced ​​feature map is input into Shuffle Attention. The feature representation is dynamically adjusted through operations such as feature grouping, shuffling, feature interaction, and attention allocation to highlight the key areas of the feature map and reduce the interference of background information. The core idea of ​​this module is to run the feature extraction path in parallel and gradually fuse features of different scales, thereby improving the network's ability to detect targets in complex scenes. The module can effectively combine features at different levels through cross-layer information fusion. Especially in road detection scenarios, the appearance of objects is often interfered by background, environmental conditions, and other objects. Cross-layer fusion can help the network understand more contextual information, so that the detector can better distinguish between targets and backgrounds.

[0050] Further, step 5) specifically includes the following steps:

[0051] Step 51) Figure 5As shown in the figure, the DHU module performs upsampling operations according to different sampling strategies. When Style = LP, the input features are directly divided into low-level features and high-level features. For these two types of features, the DHU module performs dynamic upsampling operations respectively. When Style = PL, the input features are Pixel Shuffled and then hierarchically sampled. Input the current feature map X, the input features are divided into two parts: the low-level feature map x_low and the high-level feature map x_high. The offsets of the two are calculated respectively, and then dynamic upsampling is performed according to the offsets. The feature maps after the two-level sampling are fused and output by channel splicing.

[0052] For each type of feature, the DHU module upsamples it by a dynamically calculated offset. The calculation of the offset depends on the hierarchy and structure of the input features, which enables the sampling process to adaptively handle complex scenes. The dynamic sampling process is described as follows:

[0053] Suppose the input feature is I, and it is sampled by the offset Δ. The sampled feature I is expressed as:

[0054]

[0055] Among them, R represents the sampling operation, and Δ is the offset calculated dynamically by the network.

[0056] The low-level features and high-level features after dynamic upsampling are weighted and fused with weights respectively. For the feature output of each layer, the fast normalized weighted method is used to fuse the sampled low-level and high-level features. In this weighted method, the weight is a learnable parameter, and the ReLU activation function is used to ensure that the weight is between [0,1]. The weighted formula is:

[0057]

[0058] Among them, w i and w j It is a learnable weight, and its value is ensured to be between [0,1] through the ReLU activation function to ensure the stability of the fusion process.

[0059] Each layer of the DHU module outputs features P i It can be expressed as:

[0060]

[0061] Among them, w i and w j Learned through training, and Respectively represent the low-level and high-level features after dynamic sampling.

[0062] Further, step 6) specifically includes the following steps:

[0063] Step 61) The backbone feature extraction network constructed by the above steps includes a feature fusion network of a multi-scale feature adaptive fusion module and a dynamic hierarchical upsampling module and is combined with the detection head described above to form a target detection network for occluded vehicles in urban road scenes of the present invention.

[0064] Further, step 7) specifically includes the following steps:

[0065] Step 71) input the data set into the target detection network model, configure the corresponding training environment, set the corresponding training parameters, and carry out the model training task;

[0066] Step 72) The present invention is performed on an operating system version of Ubuntu 20.04, a GPU of NVIDIA RTX 4070 (16GB), a CPU of Intel(R) Core(TM) i7-13650HX, a host memory of 16GB, a programming language of Python 3.8, and training is performed based on a deep learning framework PyTorch 1.8.1;

[0067] Step 73) In the configuration of training parameters, the optimizer uses SGD, the learning rate is set to 0.01, the weight decay rate is set to 0.0005, the training round is set to 300, and the batch size is set to 8.

[0068] Further, step 8) specifically includes the following steps:

[0069] Step 81) Load the trained weight file into the vehicle target detection network, randomly select pictures from the test data set for detection, and return the labeled pictures with the position and category information of the target to be tested.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for detecting vehicles obstructing traffic on urban roads, characterized in that: The following steps are involved: Step 1: Collect and preprocess relevant datasets for urban road scene target detection: Download the urban road scene target detection dataset from a public website and divide it into training set, test set and validation set in a ratio of 8:1:1, where the detection target is a vehicle. Use labeling software to label the target detection images of the initial dataset, mark various targets in the image with specific boxes, generate corresponding annotation files, and use them as training datasets together with the original image files; Step 2, construct a hybrid attention mechanism: The hybrid attention mechanism is composed of the global attention mechanism GAM and the coordinate attention mechanism CA. First, the channel attention module in GAM extracts channel information through MaxPool and AvgPool, generates a channel attention map, and weights it with the input features to generate channel refined features. Then, the channel refined features extract spatial features through the convolution layer, and generate a spatial attention map to highlight important areas. The CA module extracts spatial information through the maximum pooling operation to effectively suppress background interference. Finally, the features are further optimized and output through Conv2D, BatchNorm and nonlinear activation. The hybrid attention mechanism dynamically adjusts the influence of GAM and CA through a learnable weight generation module DN-Weight to improve the performance of the model in target detection tasks; Step 3, build a backbone feature extraction network: use the hybrid attention mechanism built in step 2 combined with the convolutional module consisting of convolutional layer, batch normalization layer, and activation function layer, the CSP bottleneck structure composed of CSP structure and the maximum pooling module (SPPF) to build the backbone feature extraction network; Step 4, construct a multi-scale feature adaptive fusion module: The multi-scale feature adaptive fusion module consists of a variable kernel convolution with flexible adjustment of the convolution kernel size, an enhanced residual feature extraction module ERFEM combined with residual connections and deep convolution layers, and a bottleneck module combined with an attention mechanism. The variable kernel convolution includes an adaptive convolution kernel design, which dynamically adjusts the sampling position of the feature map by generating an offset, thereby effectively capturing multi-scale information; the enhanced residual feature extraction module adds the input features to the features processed by convolution, so that the network can more easily propagate gradients during training, avoiding the gradient vanishing problem and thus accelerating the convergence speed. The multi-scale feature adaptive fusion module processes features of different scales in parallel and combines the attention mechanism to perform feature fusion. The core idea of ​​this module is to run multiple feature extraction paths in parallel and gradually fuse features of different scales, thereby improving the network's ability to detect targets in complex scenes. The module can effectively combine features at different levels through cross-layer information fusion; Step 5, construct a dynamic hierarchical upsampling module: The DHU module divides the input features into low-level features and high-level features. For these two types of features, the DHU module performs dynamic upsampling operations respectively. After the input features are divided into low-level features and high-level features, the low-level features usually contain more spatial information and details, while the high-level features contain more semantic information and global background. For each type of feature, the DHU module upsamples it through a dynamically calculated offset. The calculation of the offset depends on the hierarchy and structure of the input features, which enables the sampling process to adaptively handle complex scenes. Step 6: Construct a target detection network for occluded vehicles in urban road scenes: Use the backbone feature extraction network constructed in the above steps, including the multi-scale feature adaptive fusion module, the feature fusion network of the dynamic hierarchical upsampling module, and the detection head of the yolo network to form a target detection network for occluded vehicles in urban road scenes; Step 7, training the target detection network model: using the training set to fully train the target detection network of occluded vehicles in urban road scenes composed of the feature extraction backbone network, the multi-scale feature adaptive fusion module, the dynamic hierarchical upsampling module and the prediction network, and obtain the trained network model weights; Step 8, test the target detection network model: load the trained weight file into the target detection network, randomly select pictures from the test set for detection, return the labeled pictures with the location and category information of the target to be tested, and obtain the detection accuracy and detection speed of vehicle targets in the urban road scene image dataset, as well as the overall average accuracy.

2. The target detection method for occluding vehicles in urban road scenes according to claim 1 is characterized in that: The hybrid attention mechanism constructed in step 2 processes the target image as follows: First, the extracted high-dimensional feature map F is used as the input of the hybrid attention mechanism, and then it is simultaneously sent to the global attention module and the coordinate attention module for processing. In the global attention module, the global information between channels is first modeled through a two-layer connection network to generate a channel attention map. Then, the spatial features are extracted through two convolutional layers to generate a spatial attention map. Finally, the input features are multiplied point by point with the channel attention and spatial attention to achieve joint modeling of global channel and spatial information, so as to better highlight the significant areas in the input features; in the coordinate attention module, the significant features in each direction are extracted by performing maximum pooling operations in the height direction and the width direction respectively, and the pooling results in the two directions are spliced ​​along a specific dimension, and the features are further compressed through 1×1 convolution and activation function. Then, the features in the height and width directions are processed separately, and the corresponding attention weights are generated through additional convolutional layers. These weights are multiplied point by point with the input features, so as to enhance the expressive power of the input features in the spatial dimension. The hybrid attention mechanism dynamically weights and sums the output results of the GAM and CA modules through the weight generation module. The module uses global pooling to extract the global information of the input features, generates weights through the fully connected layer, and uses softmax to ensure that the sum of the weights is 1. The weights of the GAM and CA modules are generated according to the input feature distribution, and the influence of the two modules is dynamically adjusted to make them more adaptable to the input features. The final weighted feature map is output. Through this fusion method, the hybrid attention mechanism can not only capture the global channel and spatial relationship, but also flexibly model local salient areas, significantly improving the detection capability of target vehicles in occluded scenarios.

3. The target detection method for occluding vehicles in urban road scenes according to claim 1 is characterized in that: The multi-scale feature adaptive fusion module constructed in step 4 aims to enhance the effect of feature fusion, especially to accurately perceive and express the salient areas of targets of different scales, thereby improving the detection ability of occluded vehicles. This module is optimized for the complexity of vehicle occluded target detection in urban road scenes. The multi-scale feature adaptive fusion module includes an adaptive kernel convolution (AKConv), an enhanced residual feature extraction module (ERFEM), and a bottleneck module combined with an attention mechanism. The variable kernel convolution uses adaptively designed convolution kernels and the learned offset to dynamically adjust the sampling position of the feature map, thereby more effectively capturing multi-scale information. The enhanced residual feature extraction module provides two modes: when RC = True, residual connection is used to add the input features after 1×1 convolution to the features after variable kernel convolution, which promotes easier gradient propagation, avoids the gradient vanishing problem, and accelerates model convergence; when RC = False, residual connection is not used, thereby reducing unnecessary jump connection calculations and reducing computational overhead. In the multi-scale feature adaptive fusion module, the input feature map first passes through a 1×1 convolution block and then is divided into two parts according to the channel dimension: M1 and M2. Among them, M1 is processed by N ERFEM modules and then output. Each ERFEM module is connected in parallel with M2. The value of N is set according to the model complexity. In this design, N=2. Subsequently, the output feature maps of all ERFEM modules are spliced ​​with M2, and the spliced ​​feature maps are input into the Shuffle Attention module. Shuffle Attention dynamically adjusts feature expression, highlights key area information, and reduces background interference through operations such as feature grouping, shuffling, feature interaction, and attention allocation. The core concept of this module is to gradually fuse features of different scales through parallel feature extraction paths, thereby enhancing the network's ability to detect targets in complex scenes. In addition, through cross-layer information fusion, features at different levels can be effectively combined, which is particularly suitable for road vehicle detection tasks. In this scenario, the appearance of an object is often interfered by the background, environmental conditions, and other objects. Cross-layer fusion can help the network understand the contextual information more comprehensively, so that the detector can distinguish between targets and background more accurately.

4. The target detection method for occluding vehicles in urban road scenes according to claim 1 is characterized in that: The dynamic hierarchical upsampling module constructed in step 5 introduces a hierarchical sampling mechanism to dynamically adjust between the low-level features and high-level features of the input feature map, aiming to improve the detection effect of occluded targets. The module first divides the input feature map into low-level and high-level features along the channel dimension, where the low-level features capture detail information and the high-level features capture semantic information. The module uses a learnable convolutional layer to generate spatial offsets and applies different offsets to low-level and high-level features respectively. In order to further optimize the spatial position adjustment, an optional scope module is also introduced. The module weights the offsets and ensures that the weights are within a reasonable range through the sigmoid activation function, thereby flexibly controlling the sampling contribution of different regions. Finally, the feature map that has been spatially offset and weighted is sampled to generate a new feature map with higher resolution. In the implementation process, the module supports two different upsampling styles. In the 'LP' mode, the low-level and high-level features are sampled separately, and then fused into a unified output feature map through a splicing operation; in the 'PL' mode, the input features are first pixelated. After shuffle processing, the low-level and high-level features are offset and sampled respectively, and finally fused. Through this layered sampling mechanism, the dynamic hierarchical upsampling module can effectively fuse features of different scales while avoiding information loss, thereby improving the performance of occluded vehicle detection in complex scenes.

Images