Vehicle detection method suitable for highway tunnel monitoring scene

Through the improved ResNet-18 network and feature extraction method, the problems of uneven exposure and background interference in vehicle detection in highway tunnel scenarios are solved, the detection accuracy and anti-interference ability of vehicle targets are improved, and the complexity of highway tunnels is adapted to environment.

CN119048994APending Publication Date: 2024-11-29CHONGQING UNIV

Patent Information

Application Number
CN202410883750.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-03
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing vehicle detection method faces uneven exposure and background interference problems in highway tunnel scenes, resulting in a decline in image quality and difficulty in extracting vehicle target features. The long distance of the surveillance camera leads to large changes in the target scale, which affects detection accuracy and stability. .

Method used

Adopting the improved ResNet-18 backbone network, the feature extraction capability is enhanced through small convolution kernel convolution, average pooling structure and OrthoNet attention mechanism, and learningable position coding and Dysample upsampling methods are introduced to improve the detection accuracy of vehicle targets. and anti-interference ability.

Benefits of technology

It improves the detection accuracy and anti-interference ability of vehicle targets, enhances the adaptability to small-scale targets, and improves the vehicle detection performance in highway tunnel scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048994A_ABST
    Figure CN119048994A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of intelligent traffic, and discloses a vehicle detection method suitable for a highway tunnel monitoring scene, which comprises the following steps: S1, inputting an image to be detected into a vehicle detection model, and generating a feature map of the image to be detected in a backbone network ResNet-18 after the vehicle detection model is improved; s2, performing intra-scale feature interaction AIFI and cross-scale feature fusion CCFM on the feature map output in the step S1, improving a hybrid encoder, and inputting the feature map into a decoder to obtain an output detection result; s3, constructing a data set of vehicle target detection of the highway tunnel monitoring video; s4, training a vehicle detection model based on the highway tunnel vehicle detection data set constructed in the step S3; and S5, applying the vehicle detection model trained in the step S4 to an expressway monitoring system, and accurately detecting a vehicle target in the tunnel. The method effectively improves the detection accuracy of the small-scale vehicle target, not only better adapts to the complex environment in the expressway tunnel, but also meets the real-time requirement.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation, and in particular relates to a vehicle detection method suitable for monitoring scenes in highway tunnels. Background Art

[0002] The traffic environment in highway tunnels is very special, with narrow space and limited lighting, which brings great challenges to traffic management and safety. In order to effectively reduce traffic accidents in tunnels and ensure vehicle safety and smooth travel, it is crucial to improve the ability to perceive vehicle status. As one of the research hotspots in the field of computer vision, vehicle detection can monitor vehicle targets in real time and provide support for abnormal vehicle control, thereby promoting tunnel safety management and operational decisions, improving traffic safety and management levels, and showing broad application prospects.

[0003] However, existing vehicle detection methods still face some challenges in highway tunnel scenarios. First, there are serious exposure unevenness and background interference in the tunnel, which not only seriously affects the image quality, but also aggravates the motion blur problem of the vehicle, seriously interfering with the accurate extraction of vehicle target features. Second, due to economic costs, the distance between two equally installed surveillance cameras in the tunnel is usually far, resulting in a significant distance difference between the camera and the vehicle target, which leads to the problem of large scale changes of vehicle targets, which makes small-scale targets in the distance face the risk of being misdetected or missed. These problems have put forward severe tests on the accuracy and stability of vehicle detection algorithms. For example, Chen et al. designed a spatial pyramid structure that integrates multiple receptive fields to retain the detailed features of small targets to the greatest extent, thereby improving the detection accuracy of vehicle targets. However, this method does not take into account the uneven exposure and background interference in tunnel monitoring images, which affects the feature extraction ability of the model. Therefore, it is necessary to conduct in-depth research on a vehicle detection scheme suitable for tunnel monitoring scenarios based on the actual situation of highway tunnels. Summary of the invention

[0004] In view of this, the object of the present invention is to provide a vehicle detection method suitable for highway tunnel monitoring scenes. The present invention aims to solve the problems that the existing methods do not consider uneven exposure and background interference, have poor stability and low accuracy.

[0005] The present invention provides a vehicle detection method applicable to a highway tunnel monitoring scenario, comprising the following steps:

[0006] S1. Input the image to be detected into the vehicle detection model, and generate a feature map of the image to be detected in the improved backbone network ResNet-18 of the vehicle detection model;

[0007] S2. Perform intra-scale feature interaction AIFI and cross-scale feature fusion CCFM on the feature map output in step S1, improve the hybrid encoder, and input the feature map into the decoder to obtain the output detection result;

[0008] S3. Build a dataset for vehicle target detection in highway tunnel surveillance videos;

[0009] S4. Based on the highway tunnel vehicle detection dataset constructed in step S3, training a vehicle detection model;

[0010] S5. Apply the vehicle detection model trained in step S4 to the highway monitoring system to accurately detect vehicle targets in the tunnel.

[0011] Further, the step S1 includes the following sub-steps:

[0012] S1.1 optimizes the convolution structure;

[0013] The image to be detected is input into the backbone network ResNet-18, and the original 7×7 convolution unit in the backbone network ResNet-18 is replaced by a module consisting of two 3×3 convolutions and one 1×1 convolution in series, called VGG-ResNet;

[0014] That is, the first 3×3 convolution is used to extract primary features while reducing the spatial size of the image to retain edge information; the second 3×3 convolution is used to further deepen the feature hierarchy to ensure that the output feature map size remains unchanged; the 1×1 convolution is used to increase the number of channels to 64 to capture more complex features while maintaining the spatial size of the feature map;

[0015] S1.2 optimizes the residual structure;

[0016] The 1×1 convolution of the “short-circuit branch” in the residual blocks Stage2 to Stage4 of the backbone network ResNet-18 is replaced by a module consisting of a 1×1 convolution connected to a 2×2 average pooling layer, called AvgC-ResNet.

[0017] That is, a 1×1 convolution with a stride of 1 is used to adjust the channel dimension while keeping the spatial dimension unchanged; a 2×2 average pooling layer with a stride of 2 is used for smoother downsampling, effectively integrating the features of different channels while retaining the detail information;

[0018] S1.3 adds the OrthoNet attention module to the residual blocks Stage1 to Stage4 of ResNet-18 to perform adaptive weighting on the feature maps output by the 3×3 convolutional layer to effectively suppress and filter out non-critical information in the image, which is called Ortho-ResNet.

[0019] S1.4 Finally, the output feature map of the entire improved backbone network is obtained {S 3 ,S 4 ,S 5};

[0020] Output feature map {S 3 ,S 4 ,S 5 The downsampling times of the resolution of the input image to be detected are 8 times, 16 times and 32 times respectively.

[0021] Further, the step S2 includes the following sub-steps:

[0022] S2.1 will S 5 The feature vector generated after Flatten is added to the position code and used as the input of AIFI. The original sine-cosine position code is replaced by a learnable position code, which is called the LPE-AIFI module.

[0023] The learnable positional encoding is represented as:

[0024] PE L =Random()

[0025]

[0026] Where, PE L Represents the positional encoding initialized by the random function Random, Indicates PE L Position encoding after adaptive learning in AIFI;

[0027] S2.2 replaces the original nearest neighbor interpolation method in CCFM with Dysample, which is called Dy-CCFM module;

[0028] O = 0.5sigmoid (linear 1 (X)).linear 2 (X)

[0029] S=G+O

[0030] X' = ​​grid.sample(X,S)

[0031] Where X represents the input feature map; linear 1 and linear 2 Represent two linear transformations respectively; sigmoid represents the Sigmoid activation function; O represents the dynamically generated offset; G and S represent the original and final sampling sets respectively; X' represents that X and S are generated through the grid sampling function to generate a feature map that combines semantic and spatial information;

[0032] S2.3 inputs the feature map into the decoder for target positioning and category prediction, and finally outputs the detection result.

[0033] Further, the step S3 includes the following sub-steps:

[0034] S3.1 collects surveillance videos in highway tunnels, extracts them into frames, and labels them using the LabelImg tool;

[0035] Among them, the annotations include three categories: "car", "truck" and "bus";

[0036] S3.2 organizes the labeled vehicle detection dataset and randomly divides it into training set, test set and validation set in a ratio of 7:2:1.

[0037] Furthermore, in step S4, the training of the vehicle detection model includes the following settings:

[0038] I. Use AdamW as the optimizer, and set the initial learning rate to 0.0001, the final learning rate to 1.0, the momentum to 0.9, and the weight decay to 0.0001;

[0039] II. Using the warm-up strategy, the number of warm-up iterations is set to 2000, the warm-up momentum is set to 0.8, and the warm-up bias learning rate is set to 0.1;

[0040] III. The resolution of the input image to be detected is fixed to 640×640, the total number of training rounds is 100, and 16 images are used for training in each batch;

[0041] IV. The optimizer obtains and saves the model weights after optimizing the model loss.

[0042] Further, the step S5 includes the following sub-steps:

[0043] S5.1 loads the model weight file into the vehicle detection model trained in step S4;

[0044] S5.2 performs frame-by-frame vehicle multi-target tracking on the video stream data input into the tunnel.

[0045] Beneficial effects:

[0046] In the complex environment of the tunnel, the monitoring image is mottled due to exposure and noise interference. The ResNet-18 structure is relatively simple and it is difficult to effectively capture the local and detailed features of the vehicle target. To solve this problem, the present invention introduces small convolution kernel convolution, average pooling structure and OrthoNet attention mechanism, and designs a more robust vehicle target feature extraction network En-ResNet, which can obtain more comprehensive and detailed vehicle features, thereby enhancing the feature expression ability, small-scale target adaptability and anti-interference ability of RT-DETR.

[0047] Secondly, the fixed and uniform position coding of sine-cosine position coding cannot fully and accurately capture the dynamic changes of vehicle position (such as driving on a curve, changing lanes, etc.), and cannot solve the complex spatial relationship problems caused by the change of vehicle target position. To solve this problem, in order to better capture the dynamic characteristics of vehicle position changing with time and environment, the present invention introduces learnable position coding, which better reflects the actual position of vehicle targets in the tunnel environment through adaptive adjustment, improves the positioning accuracy of dynamically changing vehicle targets, and further improves the detection accuracy of RT-DETR for vehicle targets.

[0048] In addition, the nearest neighbor interpolation method that directly uses the value of the nearest pixel in space to fill the undetermined position is simple, but it is not sensitive enough when processing important details of vehicle targets (such as vehicle contours, etc.), especially in the scene of highway tunnels. Due to frequent occlusion between vehicles, the boundaries between different vehicles and between vehicles and backgrounds are easily blurred, affecting the positioning accuracy of the prediction frame. To solve this problem, the present invention introduces the Dysample upsampling method of adaptive learning context information, which effectively integrates the spatial and semantic information in the high-level feature map, enhances the learning ability of key features of vehicle targets, reduces the loss of important detail information, and can improve the vehicle detection performance of RT-DETR in highway tunnel scenarios.

[0049] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of a vehicle detection method applicable to a highway tunnel monitoring scenario according to the present invention;

[0051] Figure 2 For the improved backbone network;

[0052] Figure 3 It is the optimized convolution structure;

[0053] Figure 4 is the optimized residual structure;

[0054] Figure 5 This is the structural diagram of the Ortho-ResNet module;

[0055] Figure 6 Adding structural representation to positional encoding;

[0056] Figure 7 It is a display of cross-scale fusion and output structure. DETAILED DESCRIPTION

[0057] In order to make the technical solutions, advantages and purposes of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of this application.

[0058] The detection method of the present invention takes the actual highway tunnel environment as the starting point. Based on RT-DETR as the benchmark algorithm, a series of optimization methods are designed to successfully achieve accurate detection of vehicle targets in the tunnel. In view of the problem that the backbone network and efficient hybrid encoder of RT-DETR are not robust enough in complex tunnel scenes, the present invention uses ResNet-18 as the basic backbone, introduces small convolution kernel convolution, average pooling structure and OrthoNet attention mechanism to enhance the network's feature expression ability, small-scale target adaptability and noise filtering ability, and designs a more robust target feature extraction network En-ResNet. By introducing learnable position encoding and Dysample modules into the efficient hybrid encoder, the positioning accuracy of the network and the ability to capture important features are improved. This method effectively improves the detection accuracy of small-scale vehicle targets, not only better adapts to the complex environment in highway tunnels, but also meets the real-time requirements.

[0059] like Figure 1 As shown, the present invention provides a vehicle detection method suitable for highway tunnel monitoring scenarios, comprising the following steps:

[0060] S1. Input the image to be detected into the vehicle detection model, and generate the feature map of the image to be detected in the improved backbone network ResNet-18 of the vehicle detection model. Figure 2 As shown, it specifically includes the following sub-steps:

[0061] S1.1 optimizes the convolution structure;

[0062] The image to be detected is input into the backbone network ResNet-18, and the original 7×7 convolution unit in the backbone network ResNet-18 is replaced by a module consisting of two 3×3 convolutions and one 1×1 convolution in series, called VGG-ResNet;

[0063] That is, an image of size W×H×3 is input into the first 3×3 convolution, and the step size is set to 2 (s=2) to extract primary features while reducing the spatial size of the image, and the padding is set to 1 (p=1) to retain edge information; the W / 2×H / 2×3 feature map output by the first 3×3 convolution is input into the second 3×3 convolution to further deepen the feature hierarchy, and the step size and padding are both set to 1 (s=1, p=1) to ensure that the output feature map size remains unchanged; the 1×1 convolution is used to increase the number of channels by 64, allowing the model to capture more complex features by expanding the capacity of the network. At the same time, in order to maintain the spatial size of the feature map, the step size is set to 1 and the padding is set to 0 (s=1, p=0), as shown in Figure 3 As shown; among them, the network structure of ResNet-18 is shown in Table 1.

[0064] S1.2 optimizes the residual structure;

[0065] The 1×1 convolution of the “short-circuit branch” in some residual blocks (Stage2 to Stage4) of the backbone network ResNet-18 is replaced by a module consisting of a 1×1 convolution and a 2×2 average pooling layer, called AvgC-ResNet.

[0066] That is, a 1×1 convolution with a stride of 1 (s=1, p=0) is used to adjust the channel dimension and keep the spatial dimension unchanged; a 2×2 average pooling layer with a stride of 2 (s=2, p=0) performs a smoother downsampling by calculating the average value of adjacent pixels in the input feature map, effectively integrating different channel features while retaining detail information, such as Figure 4 shown.

[0067] S1.3 adds the OrthoNet attention module to the residual block (Stage1~Stage4) of ResNet-18 to perform adaptive weighting on the feature map output by the 3×3 convolutional layer to effectively suppress and filter out non-critical information (i.e., noise) in the image, which is called Ortho-ResNet;

[0068] like Figure 5 As shown, the specific process of OrthoNet is summarized as follows:

[0069]

[0070] E(Z)=σ(W 2 δ(W 1 Z))

[0071]

[0072]

[0073] In the formula, K c,h,w and X c,h,w They represent the orthogonal filter and input feature map of size H×W×C, where W, H, and C represent the width, height, and number of channels of the two, respectively; F ortho (X) c represents using K to compress the cth channel of X; W 1 and W 2 represents two fully connected layers; δ and σ represent ReLU and Sigmoid activation functions respectively; A(X) represents the process of recalibrating the features of the input image and outputting the intermediate feature map Finally, the residual connection outputs the feature map with channel attention

[0074] S1.4 Finally, the output feature map of the entire improved backbone network is obtained {S 3 ,S 4 ,S 5}; Output feature map {S 3 ,S 4 ,S 5 The downsampling times of the resolution of the input image to be detected are 8 times, 16 times and 32 times respectively;

[0075] Table 1 Network structure of ResNet-18

[0076]

[0077]

[0078] S2. Perform intra-scale feature interaction AIFI and cross-scale feature fusion CCFM on the feature map output in step S1, improve the hybrid encoder, and input the feature map into the decoder to obtain the output detection result, such as Figure 6 and Figure 7 As shown;

[0079] That is, S 5 As the output feature map of Stage 4, it is input into AIFI for feature interaction within the scale; 3 ~S 4 As the output feature map of Stage2~Stage3 and the F output of AIFI 5Input them into CCFM together to achieve cross-scale feature fusion;

[0080] It includes the following sub-steps:

[0081] S2.1 will S 5 The feature vector generated after Flattening is added to the position code and used as the input of AIFI. The original sine-cosine position code is replaced by a learnable position code, which is called the LPE-AIFI module. That is, it better reflects the actual position of the vehicle target in the tunnel environment through adaptive adjustment.

[0082] The learnable positional encoding is represented as:

[0083] PE L =Random()

[0084]

[0085] Where, PE L Represents the positional encoding initialized by the random function Random, Indicates PE L Position encoding after adaptive learning in AIFI;

[0086] S2.2 replaces the original nearest neighbor interpolation method in CCFM with Dysample, which is called Dy-CCFM module;

[0087] The specific process is:

[0088] O = 0.5sigmoid (linear 1 (X)).linear 2 (X)

[0089] S=G+O

[0090] X' = ​​grid.sample(X,S)

[0091] Where X represents the input feature map; linear 1 and linear 2 Represent two linear transformations respectively; sigmoid represents the Sigmoid activation function; O represents the dynamically generated offset; G and S represent the original and final sampling sets respectively; X' represents that X and S are generated through the grid sampling function to generate a feature map that combines semantic and spatial information;

[0092] S2.3 draws on the "denoising idea" in DN-DETR and inputs the feature map into the Decoder based on the Transformer architecture for target positioning and category prediction, and finally outputs the detection results.

[0093] S3. Build a dataset for vehicle target detection in highway tunnel surveillance videos;

[0094] S3.1 collects surveillance videos in highway tunnels, extracts them into frames, and annotates them into datasets using the LabelImg tool;

[0095] The labeling work includes labeling more than 280,000 vehicle objects, with categories of "car", "truck" and "bus". In order to ensure the diversity of training data, images are selected in different tunnel scenes, including tunnel curves, straight roads and exits;

[0096] S3.2 organizes the labeled vehicle detection dataset into the YOLO format that matches the RT-DETR algorithm, and randomly divides the labeled vehicle detection dataset into three parts in a ratio of 7:2:1: training set, test set, and validation set, which contain 42878, 12251, and 6162 surveillance images respectively.

[0097] S4. Based on the highway tunnel vehicle detection dataset constructed in step S3, training a vehicle detection model;

[0098] The training of the vehicle detection model includes the following settings:

[0099] I. Use AdamW as the optimizer, and set the initial learning rate to 0.0001, the final learning rate to 1.0, the momentum to 0.9, and the weight decay to 0.0001;

[0100] II. Using the warm-up strategy, the number of warm-up iterations is set to 2000, the warm-up momentum is set to 0.8, and the warm-up bias learning rate is set to 0.1;

[0101] III. The resolution of the input image to be detected is fixed to 640×640, the total number of training rounds is 100, and 16 images are used for training in each batch;

[0102] IV. The optimizer obtains and saves the model weights after optimizing the model loss.

[0103] S5. Apply the vehicle detection model trained in step S4 to the highway monitoring system to accurately detect vehicle targets in the tunnel.

[0104] S5.1 loads the model weight file into the vehicle detection model trained in step S4;

[0105] S5.2 performs frame-by-frame vehicle multi-target tracking on the video stream data input into the tunnel.

[0106] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should be included in the scope of the claims of the present invention.

Claims

1. A vehicle detection method suitable for highway tunnel monitoring scenarios, characterized in that: The following steps are involved: S1. Input the image to be detected into the vehicle detection model, and generate a feature map of the image to be detected in the improved backbone network ResNet-18 of the vehicle detection model; S2. Perform intra-scale feature interaction AIFI and cross-scale feature fusion CCFM on the feature map output in step S1, improve the hybrid encoder, and input the feature map into the decoder to obtain the output detection result; S3. Build a dataset for vehicle target detection in highway tunnel surveillance videos; S4. Training a vehicle detection model based on the highway tunnel vehicle detection dataset constructed in step S3; S5. Apply the vehicle detection model trained in step S4 to the highway monitoring system to accurately detect vehicle targets in the tunnel.

2. A vehicle detection method suitable for highway tunnel monitoring scenarios according to claim 1, characterized in that: The step S1 comprises the following sub-steps: S1.1 optimizes the convolution structure; The image to be detected is input into the backbone network ResNet-18, and the original 7×7 convolution unit in the backbone network ResNet-18 is replaced by a module consisting of two 3×3 convolutions and one 1×1 convolution in series, called VGG-ResNet; That is, the first 3×3 convolution is used to extract primary features while reducing the spatial size of the image to retain edge information; the second 3×3 convolution is used to further deepen the feature hierarchy to ensure that the output feature map size remains unchanged; the 1×1 convolution is used to increase the number of channels to 64 to capture more complex features while maintaining the spatial size of the feature map; S1.2 optimizes the residual structure; The 1×1 convolution of the "short-circuit branch" in the backbone network ResNet-18 residual blocks Stage2 to Stage4 is replaced by a module consisting of a 1×1 convolution connected to a 2×2 average pooling layer, called AvgC-ResNet; That is, a 1×1 convolution with a stride of 1 is used to adjust the channel dimension while keeping the spatial dimension unchanged; a 2×2 average pooling layer with a stride of 2 is used for smoother downsampling, effectively integrating the features of different channels while retaining the detail information; S1.3 adds the OrthoNet attention module to the residual blocks Stage1 to Stage4 of ResNet-18 to perform adaptive weighting on the feature maps output by the 3×3 convolutional layer to effectively suppress and filter out non-critical information in the image, which is called Ortho-ResNet. S1.4 finally obtains the output feature map {S3, S4, S5} of the entire improved backbone network; The output feature maps {S3, S4, S5} are downsampled by 8, 16, and 32 times the resolution of the input image to be detected, respectively.

3. A vehicle detection method suitable for highway tunnel monitoring scenarios according to claim 2, characterized in that: The step S2 comprises the following sub-steps: S2.1 adds the feature vector generated by Flatten in S5 to the position code and uses them as the input of AIFI. It uses the learnable position code to replace the original sine-cosine position code, which is called the LPE-AIFI module. The learnable positional encoding is represented as: ON L =Random() Where, PE L Represents the positional encoding initialized by the random function Random, Indicates PE L Position encoding after adaptive learning in AIFI; S2.2 replaces the original nearest neighbor interpolation method in CCFM with Dysample, which is called Dy-CCFM module; O=0.5sigmoid(linear1(X)).linear2(X) S=G+O X' = ​​grid.sample(X,S) Where X represents the input feature map; linear1 and linear2 represent two linear transformations respectively; sigmoid represents the Sigmoid activation function; O represents the dynamically generated offset; G and S represent the original and final sampling sets respectively; X' represents that X and S are generated through the grid sampling function to generate a feature map that combines semantic and spatial information; S2.3 inputs the feature map into the decoder for target positioning and category prediction, and finally outputs the detection result.

4. A vehicle detection method suitable for highway tunnel monitoring scenarios according to claim 3, characterized in that: The step S3 comprises the following sub-steps: S3.1 collects surveillance videos in highway tunnels, extracts them into frames, and labels them using the LabelImg tool; Among them, the annotations include three categories: "car", "truck" and "bus"; S3.2 organizes the labeled vehicle detection dataset and randomly divides it into training set, test set and validation set in a ratio of 7:2:

1.

5. A vehicle detection method suitable for highway tunnel monitoring scenarios according to claim 4, characterized in that: In step S4, the training of the vehicle detection model includes the following settings: I. Use AdamW as the optimizer, and set the initial learning rate to 0.0001, the final learning rate to 1.0, the momentum to 0.9, and the weight decay to 0.0001; II. Using the warm-up strategy, the number of warm-up iterations is set to 2000, the warm-up momentum is set to 0.8, and the warm-up bias learning rate is set to 0.1; III. The resolution of the input image to be detected is fixed to 640×640, the total number of training rounds is 100, and 16 images are used for training in each batch; IV. The optimizer obtains and saves the model weights after optimizing the model loss.

6. A vehicle detection method suitable for highway tunnel monitoring scenarios according to claim 5, characterized in that: The step S5 comprises the following sub-steps: S5.1 loads the model weight file into the vehicle detection model trained in step S4; S5.2 performs frame-by-frame vehicle multi-target tracking on the video stream data input into the tunnel.

Citation Information

Patent Citations

  • Vehicle target detection method suitable for night highway monitoring scene

    CN116883956A

  • Parking space distribution method, system and equipment based on video recognition and storage medium

    CN117237854A

  • Traffic sign detection method based on improved RT-DETR

    CN117975418A

  • Method for re-recognizing object image based on multi-feature information capture and correlation analysis

    US20220415027A1

Cited By

  • Complex scene-oriented ship number identification method and system

    CN119672694A

  • A ship number recognition method and system for complex scenes

    CN119672694B

  • In-tunnel vehicle positioning and trajectory tracking method, device and system

    CN121504991A

  • Tunnel vehicle positioning and trajectory tracking method, device and system

    CN121504991B