Vehicle detection method, system and device and storage medium
By replacing the backbone network of the YOLOv7-tiny model with ShuffleNetV2 and adding a maximum pooling structure, the problem of low accuracy and high computing resource consumption when detecting small target vehicles is solved, and efficient and accurate vehicle detection is achieved. Suitable for embedded devices and real-time systems.
Patent Information
- Application Number
- CN202411270503.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing vehicle target monitoring algorithms have low accuracy and high computing resources when detecting small target vehicles, making it difficult to deploy on embedded devices or real-time systems.
Replace the backbone network of the YOLOv7-tiny model with ShuffleNetV2, and add multiple connected maximum pooling structures and target output layers of 4×4 receptive fields to the neck layer to build a vehicle detection model, and use ShuffleNetV2 to extract feature maps and perform Feature fusion to improve detection accuracy.
It significantly reduces calculation costs, improves the detection accuracy and speed of small-target vehicles, reduces missed detection and false detection rates, and is suitable for real-time vehicle monitoring systems.
Smart Images

Figure CN119027893A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of target detection, and in particular relates to a vehicle detection method, system, device and storage medium. Background Art
[0002] With the continuous advancement of smart city construction, urban transportation systems are facing unprecedented challenges. The continuous growth of vehicle ownership has led to increasingly serious traffic congestion, pollution and safety hazards. In order to meet these challenges, people's demand for the intelligent transformation of urban transportation is becoming increasingly urgent.
[0003] As a key technology in the field of intelligent vehicle system research and intelligent transportation system, vehicle target monitoring has important research value and practical significance. Through advanced sensors and algorithms, it can detect, identify and track vehicles on the road in real time, providing key data for traffic management, traffic safety and traffic planning.
[0004] In the construction of smart cities, vehicle target monitoring technology plays a vital role. It can help city managers grasp traffic conditions in real time, detect and handle traffic incidents in a timely manner, optimize traffic signal timing, and improve traffic efficiency. In addition, it can also provide law enforcement agencies with evidence of vehicle violations and enhance traffic safety.
[0005] Existing vehicle target monitoring methods mainly include traditional algorithms and deep learning-based vehicle target monitoring algorithms. Most traditional algorithms are based on sliding windows and manually designed complex feature representations to achieve detection tasks. In complex scenarios, they have the disadvantages of high computational complexity and poor robustness. Deep learning-based algorithms have the advantages of high accuracy, fast speed, and strong robustness under complex conditions. At present, deep learning-based vehicle target monitoring algorithms are mainly divided into two categories: two-stage detection algorithms based on candidate boxes and single-stage detection algorithms based on regression. Among them, the YOLO (You Only Look Once) series of algorithms have been widely used in the field of target detection due to their advantages such as fast speed and high accuracy. However, the YOLO series of algorithms still have some urgent problems to be solved when dealing with small target vehicle detection: (1) Due to the small number of feature information of small target vehicles, it is difficult for existing vehicle monitoring algorithms to accurately detect these targets; (2) In the face of complex network structures, existing vehicle monitoring algorithms require a lot of computing resources and high-cost model training, which makes these algorithms difficult to deploy on embedded devices or real-time systems. Summary of the invention
[0006] In order to overcome the shortcomings of low detection accuracy and low detection efficiency in the prior art when detecting a vehicle, the present invention provides a vehicle detection method, comprising the following steps:
[0007] Acquire an image to be detected containing a vehicle;
[0008] The backbone network of the YOLOv7-tiny model is replaced with ShuffleNetV2, multiple sequentially connected maximum pooling structures are added to the neck layer of the YOLOv7-tiny model, and multiple target output layers with a receptive field of 4×4 are added to the prediction layer to obtain the constructed vehicle detection model;
[0009] The image to be detected is input into the vehicle detection model, ShuffleNetV2 is used to extract the feature map of the vehicle image from the image to be detected, multiple maximum pooling structures connected in sequence are used to fuse the feature map, and the target output layer is used to output the fused image to obtain the vehicle image.
[0010] Preferably, before inputting the image to be detected into the vehicle detection model, the method further includes the following steps:
[0011] Get the vehicle dataset;
[0012] The vehicle dataset is divided into training set and test set in a ratio of 9:1;
[0013] Input the images in the training set into the vehicle detection model, iteratively optimize the network parameters in the vehicle detection model, and obtain a trained vehicle detection model;
[0014] Input the images in the test set into the trained vehicle detection model and analyze the performance of the vehicle detection model.
[0015] Preferably, before inputting the images in the training set into the vehicle detection model, the method further includes preprocessing the images in the vehicle data set, specifically, performing cropping, scaling, normalization and data enhancement operations on the images.
[0016] Preferably, the ShuffleNetV2 includes multiple ShuffleNet units, each ShuffleNet unit includes multiple ShuffleNet blocks, each ShuffleNet block includes a shuffle layer, a point-by-point convolution layer and a reorganization layer, the shuffle layer is used to shuffle the channels of the feature map, the point-by-point convolution layer is used to perform nonlinear transformation on the feature map, and the reorganization layer is used to regroup the feature map after the shuffle operation.
[0017] Preferably, the ShuffleNetV2 is a staged architecture, with each stage having a different depth and width.
[0018] The present invention also provides a vehicle detection system, comprising:
[0019] An image acquisition module, used to acquire an image to be detected containing a vehicle;
[0020] The model building module is used to replace the backbone network of the YOLOv7-tiny model with ShuffleNetV2, add multiple sequentially connected maximum pooling structures to the neck layer of the YOLOv7-tiny model, and add multiple target output layers with a receptive field of 4×4 to the prediction layer to obtain the constructed vehicle detection model;
[0021] The vehicle image detection module is used to input the image to be detected into the vehicle detection model, use ShuffleNetV2 to extract the feature map of the vehicle image from the image to be detected, use multiple maximum pooling structures connected in sequence to fuse the feature map, and use the target output layer to output the fused image to obtain the vehicle image.
[0022] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the vehicle detection method.
[0023] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the vehicle detection method.
[0024] The vehicle detection method, system, device and storage medium provided by the present invention have the following beneficial effects:
[0025] The vehicle detection model constructed by the present invention can significantly reduce the calculation cost without significantly affecting the accuracy by using ShuffleNetV2 as a new feature extraction network; the present invention can improve the utilization rate of the vehicle detection model for small targets and complex scene feature semantic information by performing feature fusion on feature maps through multiple sequentially connected maximum pooling structures, thereby improving the model accuracy; by adding multiple 4×4 receptive fields, the attention paid to small target vehicles can be increased; the vehicle detection model constructed by the present invention can improve the detection accuracy of small target vehicles and reduce the missed detection and false detection rates while maintaining the speed and lightweight performance advantages of the YOLOv7-tiny model, thereby providing more effective technical support for vehicle monitoring. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. 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 these drawings without creative work.
[0027] Figure 1 The figure is a flow chart of a vehicle detection method according to an embodiment of the present invention. DETAILED DESCRIPTION
[0028] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.
[0029] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the technical solutions of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.
[0030] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.
[0031] Example 1
[0032] The present invention provides a vehicle detection method, specifically as follows Figure 1 As shown, the following steps are included:
[0033] Step 1: Get an image containing a vehicle to be detected.
[0034] Step 2: Replace the backbone network of the YOLOv7 model with ShuffleNetV2, add multiple sequentially connected maximum pooling structures to the neck layer of the YOLOv7-tiny model, and add multiple target output layers with a receptive field of 4×4 to the prediction layer to obtain the constructed vehicle detection model.
[0035] The network structure of the existing target detection YOLOv7-tiny (You Only Look Once version 7-tiny) model mainly consists of four parts, namely Input, Backbone, Neck, and Prediction. The input size of the training image is 640×640×3.
[0036] The input (input end) will use Mosaic data enhancement on the input vehicle picture, that is, read in four pictures at the same time, randomly scale, crop and arrange them, and then splice them together. Combining the four different information, it not only enriches the data set and improves the robustness of the model, but also improves the network's detection performance for small target objects due to the addition of many small targets due to random scaling, thus meeting the needs of the present invention for small target vehicle detection. In addition, according to different training sets, the optimal starting anchor frame is set by adaptive anchor frame calculation, which is convenient for iterative optimization of network parameters during training. In addition, the original picture is preprocessed to a size of 640×640×3 by using adaptive picture scaling. Compared with traditional adjustment operations, this method can effectively reduce redundant information and improve the model reasoning speed.
[0037] Backbone is the feature extraction network, which is mainly composed of three parts, namely the CBS (Convolution-BatchNorm-SiLU) module, the layer aggregation network ELAN (Efficient LayerAggregationNetwork) module and the maximum pooling MP (Max-Pooling) module. The CBS module is a combination of the convolution layer, the batch normalization layer and the SiLU activation function. YOLOv7-tiny first uses the CBS module to extract network features, perform downsampling and adjust the number of channels. In addition, the ELAN module is an efficient network structure that enables the network to learn more features and has stronger robustness by controlling the shortest and longest gradient paths. The ELAN module has two branches. The first branch changes the number of channels through a 1×1 convolution. The second branch first changes the number of channels through a 1×1 convolution module, and then passes through four 3×3 convolution modules for feature extraction. The MP module has two branches. The first branch first passes through a maximum pooling layer, which extracts the feature (maximum) data of the specified window and significantly reduces the size of the feature map (feature tensor), and then changes the number of channels through a 1×1 convolution. The second branch first changes the number of channels through a 1×1 convolution, and then passes through a 3×3 convolution kernel and a convolution block with a stride of 2. This convolution block is also used for downsampling. Finally, the results of the first and second branches are added together to obtain the super downsampling result.
[0038] Neck (neck layer) is located between Backbone and Prediction, and is composed of Feature Pyramid Network (FPN) and Path Aggregation Network (PANet). FPN is a top-down propagation path, which first performs a 2x upsampling operation on the small-sized feature map, and then performs a splicing and fusion operation with the horizontally connected feature map of the same size, and performs a 3×3 convolution operation on the fusion result to eliminate the aliasing effect caused by upsampling, and repeats the operation stage by stage to pass the strong semantic information of the deep layer to the shallow layer. PANet uses a bottom-up propagation path. It first downsamples the large-size feature map by a factor of 2, then concatenates and fuses it with the horizontally connected feature map of the same size and performs a convolution operation. It performs the same operation on the next size one by one, thereby transmitting the strong positioning information of the shallow layer to the deep layer, further improving the network's ability to extract and fuse features, improving the diversity and robustness of features, and enhancing the model's detection performance for objects of different scales. It also uses a cascade of multiple small-size pooling kernels to replace the single large-size pooling kernel in the spatial pyramid pooling SPP (Spatial Pyramid Pooling) module, which not only improves the network's recognition rate for weak targets under complex backgrounds, but also further improves the network's computing speed.
[0039] Prediction (prediction layer) is the output of the network, which is used to predict and regress the target in order to accurately locate the position of the target and predict the type of the target. YOLOv7-tiny contains three prediction layers with dimensions of (80×80×255), (40×40×255) and (20×20×255), which are used to detect large-scale targets, medium-scale targets and small-scale targets respectively. In addition, when filtering the prediction box, the weighted non-maximum suppression method is used to improve the detection effect of overlapping targets without increasing computing resources.
[0040] In summary, since the feature extraction network layer in the existing YOLOv7-tiny network structure contains a large number of convolutional layers and a more complex network structure, it is difficult to deploy in scenarios where storage space and power consumption devices are limited. Therefore, in order to meet the lightweight requirements of the model, the present invention uses a lightweight shuffle network ShuffleNetV2 (ShuffleNetworkVersion2) to replace the backbone network of YOLOv7-tiny to extract feature information of the target area. ShuffleNetV2 is a new generation of ShuffleNet version based on complementary search technology and novel architecture design, and its performance and speed are very excellent. The advantage of ShuffleNetV2 is that it has channel rearrangement and group convolution operations. Channel rearrangement is the key operation of ShuffleNetV2. Through this operation, the network can improve the efficiency of information transmission without introducing significant computational complexity, so that the channels of the feature map can be reorganized, so that information between different groups can communicate with each other, thereby better utilizing the correlation between features, further enhancing the representation ability of the model, reducing the complexity of calculation, improving the computational efficiency of the model, and realizing efficient transmission of information. The group convolution operation divides the input feature map into several groups, performs a convolution operation on each group, and finally concatenates the results. This design can reduce the amount of calculation while increasing the parallelism of the model.
[0041] In terms of structure, ShuffleNetV2 consists of multiple shuffled ShuffleNet units. Each ShuffleNet unit includes a series of ShuffleNet blocks, and each ShuffleNet block consists of a shuffle layer, a point-by-point convolution layer, and a reorganization layer. The shuffle layer implements the shuffle operation of the channel, the point-by-point convolution layer is responsible for nonlinear transformation, and the reorganization layer regroups the shuffled feature maps. By repeatedly stacking these blocks, ShuffleNetV2 is able to build a very deep network while maintaining fewer parameters and computation. In addition, ShuffleNetV2 adopts a staged architecture, with each stage having different depths and widths, which effectively controls the model, thereby reducing the number of parameters of the model while maintaining performance. In addition to the above core ideas and structures, ShuffleNetV2 also introduces some optimization strategies to improve the performance of the model, including channel splitting, point-by-point group convolution and residual connection. Channel splitting divides the input feature map into different groups, reducing the amount of calculation and communication overhead; point-by-point group convolution combines point-by-point convolution and group convolution to further reduce the amount of calculation; residual connection helps to train deeper networks and improve the convergence and generalization ability of the model.
[0042] The neck layer structure of the existing YOLOv7-tiny model also includes a spatial pyramid pooling cross-stage partial network SPPCSPC (Spatial Pyramid Pooling Cross Stage Partial Networks) module. The role of SPP is to increase the receptive field so that the YOLOv7-tiny model can adapt to images of different resolutions. It obtains different receptive fields through maximum pooling. In the first branch, there are four branches that pass through the maximum pooling layer, namely 1×1 receptive field, 5×5 receptive field, 9×9 receptive field and 13×13 receptive field. Through these four branches, convolution can process different objects. In other words, these four maximum pooling layers of different scales have four receptive fields to distinguish large targets from small targets. In addition, the SPPCSPC module divides the features into two parts, performs conventional processing on one part, performs SPP structure processing on the other part, and finally merges the two parts together, which can reduce the amount of calculation by half. This module can make the speed faster and improve the accuracy of the model, but the effect of improving the global feature fusion capability is not ideal.
[0043] The neck layer structure of the present invention adds multiple maximum pooling structures connected in sequence on the basis of SPPCSPC to obtain an improved SPPFCSPC network, which divides the image into multiple sub-regions, and then performs a maximum pooling operation in each sub-region, and finally connects the maximum pooling results of all sub-regions to form a final feature map. SPPF can better retain the global and local information of the image, thereby improving the accuracy of the target detection algorithm.
[0044] In order to increase the attention to small target vehicles and improve the detection performance, the present invention adds a tiny target prediction layer with a receptive field of 4×4 on the basis of the large, medium and small prediction layers of the existing YOLOv7-tiny model, as shown in Table 1, and redesigns the corresponding feature fusion network.
[0045] Table 1 Detection feature map information
[0046] Feature Map Prediction layer receptive field Initial prediction box size setting 160×160×64 4×4 [5,6,8,14,15,11] 80×80×128 8×8 [10,13,16,30,33,23] 40×40×256 16×16 [30,61,62,45,59,119] 20×20×512 32×32 [116,90,156,198,373,326]
[0047] Step 3: Input the image to be detected into the vehicle detection model, use ShuffleNetV2 to extract the feature map of the vehicle image from the image to be detected, use multiple sequentially connected maximum pooling structures to fuse the feature maps, and use the target output layer to output the fused image to obtain the vehicle image.
[0048] First, a vehicle dataset is obtained. The present invention uses the Car in the VisDrone2019-DET dataset as the vehicle dataset, totaling 8178 images, and preprocesses the images in the vehicle dataset, specifically: cropping, scaling, normalization, and data enhancement operations.
[0049] Secondly, the vehicle data set is divided into a training set and a test set, of which the training set accounts for 90%, a total of 7275 pictures, and the test set accounts for 10%, a total of 903 pictures. The original picture is preprocessed to a size of 640×640×3 by an adaptive picture scaling method. Compared with the traditional adjustment operation, this method can effectively reduce redundant information and improve the model reasoning speed. Then, the input image is subjected to a mosaic data enhancement operation, that is, four pictures are read in at the same time, and they are randomly scaled, cropped and arranged before being spliced and combined. Combining the four different information, not only enriches the data set, but also improves the robustness of the model. At the same time, since random scaling adds many small targets, the network's detection performance for small target objects is improved, meeting the requirements of the present invention for small target vehicle detection. In addition, according to different training sets, the optimal starting anchor frame is set by an adaptive anchor frame calculation method, which is convenient for iterative optimization of network parameters during training.
[0050] Finally, the image to be detected is input into the trained vehicle detection model, ShuffleNetV2 is used to extract the feature map of the vehicle image, multiple maximum pooling structures connected in sequence are used to fuse the feature map, and the target output layer is used to output the fused image to obtain the vehicle image.
[0051] The present invention also provides a vehicle detection system, including an image acquisition module, a model construction module and a vehicle image detection module. The image acquisition module is used to acquire an image to be detected containing a vehicle; the model construction module is used to replace the backbone network of the YOLOv7-tiny model with ShuffleNetV2, add multiple sequentially connected maximum pooling structures to the neck layer of the YOLOv7-tiny model, and add multiple target output layers with a receptive field of 4×4 to the prediction layer to obtain a constructed vehicle detection model; the vehicle image detection module is used to input the image to be detected into the vehicle detection model, use ShuffleNetV2 to extract a feature map of the vehicle image from the image to be detected, use multiple sequentially connected maximum pooling structures to perform feature fusion on the feature map, and use the target output layer to output the fused image to obtain a vehicle image.
[0052] The present invention also provides a computer device, comprising a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the vehicle detection method.
[0053] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is suitable for being loaded by a processor to execute the vehicle detection method.
[0054] Example 2
[0055] In order to accurately evaluate the performance of the vehicle detection model, this paper analyzes the model lightweight and detection accuracy based on the VisDrone2019-DET dataset, and selects Parameters as an important core indicator for measuring the consumption of computing memory resources. In addition, it also includes Recall, MAP@0.5 (average precision when the threshold intersection ratio is 0.5), MAP@0.5:0.95 (with 0.05 as the step size, the average precision of the threshold intersection ratio from 0.5 to 0.95), GFLOPS (Giga Floating-point Operations Per Second, 1 billion floating-point operations per second), frame rate / FPS (the number of images that can be detected per second) and Model Size (model size), a total of 7 different performance indicators, to measure the performance of the model from multiple aspects and angles. The calculation formula is as follows:
[0056]
[0057] GFLOPS = 2 × 10 -9 HW(C in K 2 +1)C out
[0058]
[0059] In the above formula, three binary classification parameters are publicly selected, and their definitions are shown in Table 2, where 1 and 0 respectively indicate whether the result is a vehicle, TP is a true positive sample, indicating the number of correctly detected vehicles, that is, the number of samples that are actually vehicles and predicted by the model as vehicles, and FP is a false positive sample, indicating the number of false positive samples of vehicles.
[0060] Table 2 Definition of binary classification parameters
[0061] parameter True value Predicted value TP 1 1 FP 0 1 FN 1 0
[0062] FN is the number of false negative samples, which indicates the number of missed vehicle detections, that is, the number of samples that are actually vehicles but predicted by the model as non-vehicles. AP is the area under the precision-recall curve. MAP is the average precision. GFLOPS is the number of billion floating-point operations per second. n is the total number of categories. HW is the size of the output feature map. K is the size of the convolution kernel. C in With C out are the number of input channels and output channels respectively, FPS is the number of images that can be detected per second, and t is the time consumed to detect one image. For the convenience of description, The P in it stands for Precision and the R stands for Recall.
[0063] The recall rate of the vehicle detection model of the present invention was finally improved by 1.72%, the average detection accuracy increased by 4.14%, and 55 pictures can be detected per second, meeting the real-time requirements. Compared with the YOLOv7 algorithm and mainstream vehicle detection algorithms, the improved algorithm of the present invention has higher detection accuracy and faster detection speed, and is more suitable for real-time detection of small target vehicles under edge computing.
[0064] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.
Claims
1. A vehicle detection method, characterized in that: The steps include: Acquire an image to be detected containing a vehicle; The backbone network of the YOLOv7-tiny model is replaced with ShuffleNetV2, multiple sequentially connected maximum pooling structures are added to the neck layer of the YOLOv7-tiny model, and multiple target output layers with a receptive field of 4×4 are added to the prediction layer to obtain the constructed vehicle detection model; The image to be detected is input into the vehicle detection model, ShuffleNetV2 is used to extract the feature map of the vehicle image from the image to be detected, multiple maximum pooling structures connected in sequence are used to fuse the feature map, and the target output layer is used to output the fused image to obtain the vehicle image.
2. The vehicle detection method according to claim 1, characterized in that: Before inputting the image to be detected into the vehicle detection model, the following steps are also included: Get the vehicle dataset; The vehicle dataset is divided into training set and test set in a ratio of 9:1; Input the images in the training set into the vehicle detection model, iteratively optimize the network parameters in the vehicle detection model, and obtain a trained vehicle detection model; Input the images in the test set into the trained vehicle detection model and analyze the performance of the vehicle detection model.
3. The vehicle detection method according to claim 2, characterized in that: Before inputting the images in the training set into the vehicle detection model, the method also includes preprocessing the images in the vehicle data set, specifically: cropping, scaling, normalizing and data enhancement operations on the images.
4. The vehicle detection method according to claim 1, characterized in that: The ShuffleNetV2 includes multiple ShuffleNet units, each ShuffleNet unit includes multiple ShuffleNet blocks, each ShuffleNet block includes a shuffle layer, a point-by-point convolution layer and a reorganization layer, the shuffle layer is used to shuffle the channels of the feature map, the point-by-point convolution layer is used to perform nonlinear transformation on the feature map, and the reorganization layer is used to regroup the feature map after the shuffle operation.
5. The vehicle detection method according to claim 1, characterized in that: The ShuffleNetV2 is a staged architecture, with each stage having a different depth and width.
6. A vehicle detection system, characterized in that: include: An image acquisition module, used to acquire an image to be detected containing a vehicle; The model building module is used to replace the backbone network of the YOLOv7-tiny model with ShuffleNetV2, add multiple sequentially connected maximum pooling structures to the neck layer of the YOLOv7-tiny model, and add multiple target output layers with a receptive field of 4×4 to the prediction layer to obtain the constructed vehicle detection model; The vehicle image detection module is used to input the image to be detected into the vehicle detection model, use ShuffleNetV2 to extract the feature map of the vehicle image from the image to be detected, use multiple maximum pooling structures connected in sequence to fuse the feature map, and use the target output layer to output the fused image to obtain the vehicle image.
7. A computer device, characterized in that: It comprises a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the vehicle detection method according to any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the vehicle detection method according to any one of claims 1 to 5.
Citation Information
Patent Citations
Lightweight power distribution assembly defect identification method based on improved YOLOv5-LITE
CN114862768A
Vehicle detection and identification method based on YOLOv5 model optimization
CN116311154A
Multi-scale high-precision lightweight target detection method based on large receptive field and attention mechanism
CN117058542A
YOLO-based lightweight high-precision vehicle target detection method
CN118135501A
Light-weighted target detection method and device, and storage medium
WO2022213395A1
