Vehicle detection method and system based on unmanned aerial vehicle and improved YOLO algorithm

Abstract

The invention discloses a vehicle detection method and system based on an unmanned aerial vehicle and an improved YOLO algorithm, and the method comprises the following steps: data acquisition: collecting an aerial image of the unmanned aerial vehicle; a data preprocessing step: carrying out image preprocessing operation on the acquired image data; and a target detection step of pre-building a target detection model based on a YOLO algorithm, performing target detection according to the preprocessed unmanned aerial vehicle aerial image, obtaining a target detection result, and completing vehicle detection and identification. According to the invention, by improving the target detection model, efficient and accurate detection of the vehicle can be realized under the view angle of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a vehicle detection method and system based on an unmanned aerial vehicle and an improved YOLO algorithm. Background Art

[0002] With the rapid development of drone technology, images collected by drone platforms have been widely used in various monitoring and detection tasks, especially in the key field of vehicle detection. Traditional vehicle detection methods mainly rely on fixed cameras or ground sensors. Such methods are usually limited by factors such as shooting angle, coverage, and environmental complexity, resulting in low detection accuracy and insufficient detection efficiency in large-scale and highly complex scenes. Especially in the detection scenarios of long-distance and small targets, vehicles often cause missed detection or false detection in existing detection algorithms due to their scale differences, background complexity, and changing lighting conditions.

[0003] In recent years, deep neural networks, especially object detection algorithms based on convolutional neural networks, have demonstrated high automatic feature learning capabilities and detection accuracy in vehicle detection tasks in drone images.

[0004] For example, the existing YOLO series target detection models perform well in terms of detection speed and accuracy, but in the case of multi-scale target detection, complex background processing and high real-time requirements, the model's robustness and detection accuracy still need to be improved. Therefore, it is urgent to innovate and optimize the existing target detection algorithm to better adapt to the complex requirements of drone vehicle detection scenarios and improve detection accuracy, robustness and real-time performance. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a vehicle detection method based on a drone and an improved YOLO algorithm. By improving the target detection model, the present invention can achieve efficient and accurate detection of vehicles from the perspective of a drone.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] The vehicle detection method based on drone and improved YOLO algorithm includes the following steps:

[0008] Data acquisition step, collecting drone aerial images;

[0009] A data preprocessing step, performing image preprocessing operations on the acquired image data;

[0010] In the target detection step, a target detection model based on the YOLO algorithm is built in advance, and target detection is performed according to the preprocessed drone aerial images to obtain target detection results and complete vehicle detection and identification.

[0011] The data acquisition steps are as follows: the drone is equipped with a high-resolution camera, which drives the drone to fly to capture and collect images of the target area at various heights and angles.

[0012] The image preprocessing operations in the data preprocessing step include: image enhancement, denoising and / or correction.

[0013] The target detection model is improved based on the YOLOv10 algorithm model. The improvements to its network structure based on YOLOv10 include:

[0014] Replace the BottleNeck in the C2f module in its backbone network with the GateNeck module;

[0015] The CGSFR-FPN context-guided spatial feature reconstruction pyramid network CGSFR-FPN is introduced in the Neck module of the YOLOv10 network.

[0016] The GateNeck module includes the SMPConv module, the SiLU activation function, the CGLU module and the residual connection. After the feature map is input into the GateNeck module, the SMPConv module uses a convolution kernel of a specific size to extract spatial features. The feature map after convolution processing is then nonlinearly transformed through the SiLU activation function and then sent to the CGLU module for feature enhancement and background noise suppression. The feature map processed by CGLU is residually connected with the input feature map to obtain the output of the GateNeck module.

[0017] The CGLU module includes multiple linear transformations and convolution operations. The CGLU module divides the input feature map into two paths, one through linear transformation and the other through convolution and activation processing to obtain two feature maps. The two feature maps are then gated by pixel-by-pixel multiplication to achieve enhanced expression of important features and suppression of background noise.

[0018] The convolution operation in the CGLU module uses depthwise separable convolution DW Conv to extract the spatial information of features.

[0019] The context-guided spatial feature reconstruction pyramid network CGSFR-FPN includes a pyramid context extraction module PCE, a rectangle calibration module RCM, a dynamic interpolation fusion module DIF and a multi-scale fusion module FBM; wherein the feature maps P3, P4 and P5 from the backbone network represent feature maps of different scales respectively;

[0020] The feature maps P3, P4 and P5 are sequentially input into the pyramid context extraction module to extract the important context in the features at each level and generate new multi-scale feature maps P3', P4' and P5'; the calibrated feature maps P5', P4' and P3' will further enter the multi-scale fusion module to achieve the effective combination of high-level semantic information and low-level spatial information; at the same time, the initial P3, P4 and P5 feature maps are calibrated through the rectangular calibration module to enhance the perception of the target area; the calibrated features are fused through the dynamic interpolation module to achieve high-scale feature maps that can be spatially aligned with low-scale feature maps, achieving more accurate cross-scale fusion.

[0021] The fused feature map will be sent to the target detection head to perform vehicle target classification and bounding box regression operations to complete the detection task of targets of different scales.

[0022] A vehicle detection system based on a drone and an improved YOLO algorithm, the system comprising:

[0023] A data acquisition module, which is configured to acquire drone aerial vehicle images;

[0024] A data preprocessing module, which is configured to perform image data preprocessing operations on the images taken by the drone;

[0025] The target detection module is used to perform target detection on the pre-processed drone aerial images according to the pre-built target detection model based on the YOLO algorithm, obtain the target detection results, and complete vehicle detection and identification.

[0026] The advantages of the present invention are: by introducing the GateNeck module and the context-guided spatial feature reconstruction pyramid network (CGSFR-FPN) into the YOLOv10 network, the speed, accuracy and robustness of vehicle detection are significantly improved. The model can effectively process vehicle targets in multi-angle and multi-scale images of drones, especially in complex backgrounds and different scales, to ensure accurate detection of vehicles. At the same time, the optimized network structure reduces the computational overhead and improves the real-time processing capability, so that it can adapt to various application scenarios such as drone monitoring, traffic management and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The following is a brief description of the contents expressed in the drawings of the present invention and the symbols in the drawings:

[0028] Figure 1 Flow chart of a vehicle detection method based on a drone and an improved YOLO algorithm in Embodiment 1 of the present invention;

[0029] Figure 2(a) is a schematic diagram of the structure of the CGLU module in Example 1 of the present invention;

[0030] Figure 2 (b) is a schematic diagram of the structure of the GateNeck module in the first embodiment of the present invention;

[0031] Figure 3 It is a schematic diagram of the structure of a context-guided spatial feature reconstruction pyramid network (CGSFR-FPN) in the first embodiment of the present invention;

[0032] Figure 4 The structure diagram of the vehicle detection system based on drone and improved YOLO algorithm in the second embodiment of the present invention. DETAILED DESCRIPTION

[0033] The specific implementation of the present invention will be further explained in detail below by describing the optimal embodiment with reference to the accompanying drawings.

[0034] Embodiment 1

[0035] This embodiment scheme relates to a vehicle detection method based on a drone and an improved YOLO algorithm, which aims to solve the problems of insufficient accuracy and low efficiency of traditional vehicle detection methods in large-scale and highly complex environments. By improving the target detection model, the present invention can achieve efficient and accurate detection of vehicles from the perspective of drones. Figure 1-4 As shown in the figure, a vehicle detection method based on drone and improved YOLO algorithm is proposed. It includes:

[0036] Data acquisition, obtaining drone aerial images;

[0037] Data preprocessing: performing image preprocessing operations on the acquired image data;

[0038] Target detection: Target detection is performed based on the acquired drone images and the preset target detection model to obtain target detection results and complete vehicle detection and identification.

[0039] As one or more implementation modes, this example uses a high-resolution camera carried by a drone to collect data. The camera collects multi-angle and multi-scale ground vehicle data of the target area at different flight altitudes (30 meters, 50 meters, 100 meters, etc.) and at various angles.

[0040] In order to accurately mark the targets in the acquired drone aerial images, this embodiment uses Labelimg for marking, and the marking content includes the bounding box and category information of the vehicle. The marking data adopts the YOLO format, and each marked object is described by five numerical values, namely the target category, the x-coordinate and y-coordinate of the center point of the bounding box, and the width and height of the bounding box. These values ​​are normalized.

[0041] As one or more implementation modes, this example uses multiple methods to preprocess the data of the annotated drone aerial images, including but not limited to the following operations:

[0042] Image enhancement, by adjusting image contrast, brightness and other operations, enhances the visual effect of the image to ensure that the vehicle can be clearly identified under different lighting conditions.

[0043] Image denoising, use image denoising technology to remove random noise in the image to improve image clarity and reduce interference.

[0044] Image correction: geometric correction is performed on the image to eliminate image distortion caused by changes in the flight angle of the drone and ensure the geometric consistency of the image in subsequent inspection processes.

[0045] As one or more implementation modes, this example performs target detection based on the acquired drone image and a preset target detection model, obtains a target detection result, and completes vehicle detection and identification.

[0046] In this embodiment, the target detection model used adopts the improved YOLOv10 algorithm. Specifically, the BottleNeck of the C2f module in the YOLOv10 backbone network is replaced by the GateNeck module. The GateNeck module is designed to enhance the model's feature extraction and selection capabilities under complex backgrounds. Unlike the traditional Bottleneck module, GateNeck introduces more complex convolution and gating mechanisms to effectively suppress background noise while retaining key features. By improving the Neck part of YOLOv10, a Context-Guided Spatial Feature Reconstruction FeaturePyramid Network (CGSFR-FPN) module is proposed to replace the traditional FPN+PAN structure to solve the problem of insufficient detection accuracy and robustness caused by large target scale changes and complex backgrounds in drone image vehicle detection. CGSFR-FPN achieves accurate detection of targets of different scales by introducing multi-scale context information and spatial feature reconstruction mechanisms.

[0047] The structure diagram of the GateNeck module is as follows Figure 2As shown in (b), the input feature map first enters the SMPConv module, which uses a convolution kernel of a specific size (such as 13x13) to extract spatial features and combines the method of dynamic sampling points to adapt to the diversity and uncertainty of the input features. The feature map after convolution is then nonlinearly transformed through the SiLU (Sigmoid Linear Unit) activation function. The SiLU activation function smoothes the feature map after convolution, enhances the continuity of feature expression, and effectively avoids the problem of gradient disappearance, ensuring the effective transmission of the feature map in subsequent processing.

[0048] The activated feature map further enters the CGLU (Convolutional Gated Linear Unit) module. The structure of the CGLU module is as follows Figure 2 As shown in (a), it includes multiple linear transformations and convolution operations. The CGLU module divides the input feature map into two paths, one through linear transformation and the other through convolution and activation processing. The convolution part uses deep separable convolution (DW Conv) to further extract the spatial information of the features. The two feature maps are then gated by pixel-by-pixel multiplication to achieve enhanced expression of important features and suppression of background noise. Finally, the feature map processed by CGLU is residually connected to the input feature map, that is, pixel-by-pixel addition. Through this operation, the GateNeck module not only retains the basic information of the original input, but also strengthens the key features through the gating mechanism. The output feature map after residual connection is more expressive and provides high-quality feature input for subsequent target detection tasks.

[0049] The context-guided spatial feature reconstruction pyramid network (CGSFR-FPN) in this embodiment is as follows Figure 3 As shown in the figure. This module consists of four main parts: Pyramid Context Extraction Module (PCE), Rectangular Calibration Module (RCM), Dynamic Interpolation Fusion Module (DIF) and Multi-scale Fusion Module (FBM). It aims at the detection difficulties caused by the change of flight altitude, the size of target vehicles and the complexity of background in UAV images. Through context guidance and multi-scale feature fusion, the detection accuracy of small targets is effectively improved, and false detection and missed detection are reduced in scenes with complex backgrounds. At the same time, the module's spatial feature reconstruction capability can enable the model to capture the boundary information of the vehicle more accurately and improve the detection accuracy.

[0050] The feature fusion process of CGSFR-FPN in this example is as follows: First, the feature maps P3, P4, and P5 from the backbone network represent feature maps of different scales. P3 represents a higher resolution but shallower feature map, P4 represents a medium resolution and depth feature map, and P5 represents a lower resolution but deeper feature map. The feature maps P3, P4, and P5 are sequentially input into the pyramid context extraction module to extract important contexts in the features at each level and generate new multi-scale feature maps P3', P4', and P5'. The calibrated feature maps P5', P4', and P3' will further enter the multi-scale fusion module to achieve an effective combination of high-level semantic information and low-level spatial information, thereby enhancing the expressiveness of the features. At the same time, the initial P3, P4, and P5 feature maps are calibrated through the rectangular calibration module to enhance the perception of the target area. In order to perform more refined cross-layer feature fusion in the multi-scale fusion process, the calibrated features will also be used through the dynamic interpolation fusion module to achieve spatial alignment of high-scale feature maps with low-scale feature maps, thereby achieving more accurate cross-scale fusion.

[0051] Finally, the feature map after CGSFR-FPN fusion processing will be sent to the target detection head to perform vehicle target classification and bounding box regression operations to complete the detection task of targets of different scales. This process ensures the effective fusion of features at different resolutions through the collaboration of multiple modules, improving the model's robustness and detection accuracy for complex backgrounds and multi-scale targets. The components of CGSFR-FPN are:

[0052] 1) Pyramid Context Extraction Module (PCE)

[0053] The pyramid context extraction module is responsible for extracting multi-scale context information from feature maps of different scales to ensure the perception of vehicles of different sizes and distances. This module performs convolution and adaptive pooling on multiple feature maps (P3, P4, P5) from Backbone, and then concatenates these feature maps of different scales to form a fused feature map with multi-scale context information.

[0054] In the specific implementation process of the present invention, the pyramid context extraction module first applies a convolution with a convolution kernel size of 1 and a step size of 1 to each input feature map to reduce the number of channels of the feature map and reduce the computational complexity. Then, the feature maps of different scales are unified to the same spatial size through adaptive average pooling. Finally, the feature maps of each scale are concatenated in the channel dimension to form a multi-scale context fusion feature map.

[0055] 2) Rectangular Calibration Module (RCM)

[0056] The rectangular calibration module is mainly used to finely calibrate the multi-scale feature map to reduce the interference of background noise and enhance the feature expression of the vehicle area.

[0057] In the specific implementation process of the present invention, the rectangular calibration module first applies a depth-separable convolution with a kernel size of 3, a step size of 1, and a padding of 1 to the multi-scale context feature map to reduce the amount of calculation and retain key spatial information. Then, the feature map is reorganized using a convolution with a kernel size of 1 and a step size of 1. Subsequently, the feature map is weighted and calibrated pixel by pixel through a nonlinear transformation of the ReLU activation function and a calibration weight map generated by the Sigmoid function. This process can effectively suppress background noise and highlight the characteristic area of ​​the vehicle, thereby improving detection accuracy.

[0058] 3) Dynamic interpolation fusion module (DIF)

[0059] The dynamic interpolation fusion module is used to fuse feature maps across scales to ensure the effective combination of high-scale feature maps and low-scale feature maps. By dynamically adjusting the resolution of high-scale feature maps, the module can achieve effective fusion between features of different scales.

[0060] In the specific implementation of the present invention, the dynamic interpolation fusion module first performs convolution on the high-scale feature map with a convolution kernel size of 1 and a step size of 1 to adjust its channel number to be consistent with the channel number of the low-scale feature map. Then, the spatial size of the high-scale feature map is adjusted to be consistent with the low-scale feature map through bilinear interpolation. Figure 1 Finally, the interpolated high-scale feature map is added pixel by pixel to the low-scale feature map to complete the fusion of cross-scale features and ensure that objects of different scales can be effectively detected.

[0061] 4) Multi-scale Fusion Module (FBM)

[0062] The multi-scale fusion module is responsible for further integrating the multi-scale feature maps after interpolation fusion to enhance the feature expression ability of the target. By cross-fusion of feature maps of different scales, it ensures that the spatial information of shallow features and the semantic information of deep features can be fully combined.

[0063] In the specific implementation of the present invention, two convolutions with a kernel size of 3, a stride of 1, and a padding of 1 are first applied to ensure that the spatial information and semantic information of the feature map can be effectively transmitted. Subsequently, the convolution feature map is normalized by BatchNorm, and the ReLU activation function is used to improve the nonlinear expression ability of the model, ensuring that the fused feature map can accurately represent the key features of the target vehicle.

[0064] The CGSFR-FPN module in this embodiment can significantly improve the performance of vehicle detection in drone images. Through context guidance and multi-scale feature fusion, the module performs well in the face of complex backgrounds and multi-scale targets, and is particularly suitable for the detection of long-distance and small target vehicles captured during drone flight. At the same time, the CGSFR-FPN module can effectively reduce the computational overhead, maintain the efficiency of the model, and meet the needs of real-time drone detection.

[0065] In this example, the Head part of the YOLOv10 network is responsible for object detection and classification of the multi-scale feature maps output by the Neck. After a series of convolution operations, the feature maps of each scale output the corresponding category and bounding box prediction. The Head part combines the non-maximum suppression (NMS) algorithm to remove redundant detection boxes and finally output the vehicle detection results.

[0066] Embodiment 2

[0067] Reference Figure 4 Embodiment 2 of the present invention introduces a vehicle detection system based on a drone and an improved YOLO algorithm. A vehicle detection system based on a drone and an improved YOLO algorithm includes multiple modules, which are respectively used to complete the acquisition, processing and detection of drone vehicle images, including:

[0068] A data acquisition module, which is configured to acquire drone aerial vehicle images;

[0069] a data preprocessing module configured to preprocess image data;

[0070] A target detection module is configured to perform target detection based on the acquired drone image and a preset target detection model, obtain a target detection result, and complete vehicle detection and recognition of the drone image;

[0071] Specifically, the present invention adopts an improved YOLOv10 algorithm as a target detection model. By using the GateNeck module in the Backbone part of the network, through the combination of multi-layer convolution and gating mechanism, while maintaining computational efficiency, more recognizable features can be extracted in the complex environment photographed by the drone. At the same time, the Context-Guided Spatial Feature Reconstruction Feature Pyramid Network (CGSFR-FPN) context-guided spatial feature reconstruction pyramid network is introduced in the Neck part of the YOLOv10 network. By guiding multi-scale context information, accurate spatial feature reconstruction and multi-scale feature fusion are achieved, thereby improving the accuracy and robustness of complex background and vehicle detection of different scales from the perspective of the drone.

[0072] The data acquisition module, the data preprocessing module, and the target detection module are implemented by the methods corresponding to the data acquisition step, the data preprocessing step, and the target detection step of the first embodiment, respectively. For specific steps, see the first embodiment. This scheme adopts the improved network of YOLOV10, and introduces the GateNeck module and the context-guided spatial feature reconstruction pyramid network (CGSFR-FPN) in the YOLOv10 network, which significantly improves the speed, accuracy and robustness of vehicle detection. The model can effectively process vehicle targets in multi-angle and multi-scale images of drones, especially in complex backgrounds and different scales, to ensure accurate detection of vehicles. At the same time, the optimized network structure reduces the computational overhead and improves the real-time processing capability, so that it can adapt to various application scenarios such as drone monitoring, traffic management and emergency response.

[0073] Obviously, the specific implementation of the present invention is not limited to the above-mentioned methods. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, they are all within the protection scope of the present invention.

Claims

1. A vehicle detection method based on drone and improved YOLO algorithm, characterized by: The steps include: Data acquisition step, collecting drone aerial images; A data preprocessing step, performing image preprocessing operations on the acquired image data; In the target detection step, a target detection model based on the YOLO algorithm is built in advance, and target detection is performed according to the preprocessed drone aerial images to obtain target detection results and complete vehicle detection and identification.

2. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 1, characterized in that: The data acquisition steps are as follows: the drone is equipped with a high-resolution camera, which drives the drone to fly to capture and collect images of the target area at various heights and angles.

3. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 1, characterized in that: The image preprocessing operations in the data preprocessing step include: image enhancement, denoising and / or correction.

4. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 1, characterized in that: The target detection model is improved based on the YOLOv10 algorithm model. The improvements to its network structure based on YOLOv10 include: Replace the BottleNeck in the C2f module in its backbone network with the GateNeck module; The CGSFR-FPN context-guided spatial feature reconstruction pyramid network CGSFR-FPN is introduced in the Neck module of the YOLOv10 network.

5. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 4, characterized in that: The GateNeck module includes the SMPConv module, the SiLU activation function, the CGLU module and the residual connection. After the feature map is input into the GateNeck module, the SMPConv module uses a convolution kernel of a specific size to extract spatial features. The feature map after convolution processing is then nonlinearly transformed through the SiLU activation function and then sent to the CGLU module for feature enhancement and background noise suppression. The feature map processed by CGLU is residually connected with the input feature map to obtain the output of the GateNeck module.

6. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 5, characterized in that: The CGLU module includes multiple linear transformations and convolution operations. The CGLU module divides the input feature map into two paths, one through linear transformation and the other through convolution and activation processing to obtain two feature maps. The two feature maps are then gated by pixel-by-pixel multiplication to achieve enhanced expression of important features and suppression of background noise.

7. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 6, characterized in that: The convolution operation in the CGLU module uses depthwise separable convolution DW Conv to extract the spatial information of features.

8. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 4, characterized in that: The context-guided spatial feature reconstruction pyramid network CGSFR-FPN includes a pyramid context extraction module PCE, a rectangle calibration module RCM, a dynamic interpolation fusion module DIF and a multi-scale fusion module FBM; wherein the feature maps P3, P4 and P5 from the backbone network represent feature maps of different scales respectively; The feature maps P3, P4 and P5 are sequentially input into the pyramid context extraction module to extract the important context in the features at each level and generate new multi-scale feature maps P3', P4' and P5'; the calibrated feature maps P5', P4' and P3' will further enter the multi-scale fusion module to achieve the effective combination of high-level semantic information and low-level spatial information; at the same time, the original P3, P4 and P5 feature maps are respectively calibrated through the rectangular calibration module to enhance the perception of the target area; The calibrated features are fused through a dynamic interpolation module to achieve spatial alignment of high-scale feature maps with low-scale feature maps, thus achieving more accurate cross-scale fusion.

9. The vehicle detection method based on drone and improved YOLO algorithm as claimed in claim 8, characterized in that: The fused feature map will be sent to the target detection head to perform vehicle target classification and bounding box regression operations to complete the detection task of targets of different scales.

10. A vehicle detection system based on drone and improved YOLO algorithm, characterized by: The system comprises: A data acquisition module, which is configured to acquire drone aerial vehicle images; A data preprocessing module, which is configured to perform image data preprocessing operations on the images taken by the drone; The target detection module is used to perform target detection on the pre-processed drone aerial images according to the pre-built target detection model based on the YOLO algorithm, obtain the target detection results, and complete vehicle detection and identification.

Images