Ground vehicle target detection algorithm based on dual-light aerial image of unmanned aerial vehicle

Abstract

The invention discloses an unmanned aerial vehicle dual-light aerial image over-the-ground vehicle target detection algorithm. The algorithm comprises the steps that an unmanned aerial vehicle pod collects ground visible light and thermal imaging data in real time; preprocessing and enhancing the sample; detecting a ground vehicle target; respectively extracting the features of the two images by using a trunk feature extraction network, and fusing the features at different depths, so that the vehicle target is more significant in the feature map; the feature extraction network is enhanced to further process fusion features; and inputting the processed feature map into a target detection head, predicting the type and the position of the vehicle, and transmitting a video with a mark back to a ground computer. According to the invention, for the aerial image of the unmanned aerial vehicle, the accuracy and efficiency of vehicle target detection are effectively improved, and powerful information support is provided for aerial observation and traffic management.

Description

Technical Field

[0001] The present invention relates to a target detection algorithm and belongs to the field of deep learning. Background Art

[0002] In the field of UAV ground observation, accurate detection of vehicle targets is one of the key technologies. However, single spectrum images often fail to provide sufficient information to ensure the accuracy and robustness of detection under complex lighting conditions, such as at night, in haze or shadow areas. Therefore, it is particularly important to develop a dual-light fusion target detection algorithm for visible light and infrared images. Dual-light fusion can make full use of the clear details of visible light images during the day and under good lighting conditions, as well as the thermal radiation information of thermal imaging images at night and under low light conditions, thereby achieving all-weather and high-precision detection of vehicle targets. However, the existing dual-light fusion algorithms cannot adapt to the detection of small targets in UAV images in terms of feature extraction and fusion strategy design, and cannot meet the high-precision requirements for vehicle target detection in practical applications. In view of this, a dual-light aerial image ground vehicle target detection algorithm for UAVs is proposed, aiming to improve the accuracy of ground vehicle target detection in UAV aerial images to meet the needs of practical applications.

[0003] CN112488043A discloses a moving vehicle detection method based on drone aerial video. First, the SURF algorithm is used to match feature points and remove abnormal points on the image. The drone image registration algorithm that combines global and local homography matrices is used to obtain the transformation matrix to compensate for the adverse effects of the airborne camera movement. Subsequently, the 2-frame difference method is used to reduce the area to be detected. The area to be detected is then traversed according to the center of the superpixel to further improve the efficiency of moving vehicle detection. Then, the multi-channel HOG feature algorithm is used to extract the low-order features of the vehicle, and the context information of the vehicle is introduced to obtain the high-order features of the vehicle. The two features are fused to obtain the multi-order features of the target vehicle. Finally, the multi-order features and the dictionary learning algorithm are combined to realize moving vehicle detection. This method can suppress the influence of the drone airborne camera movement, process the vehicle deformation and background interference in the image, and improve the robustness and real-time performance of moving vehicle detection.

[0004] CN112488043A provides a method for detecting unmanned aerial vehicle targets based on edge intelligence, the method comprising: deploying a two-stage filter and a lightweight target detection model S_YOLOv3 on the unmanned aerial vehicle side, and deploying a high-precision target detection model YOLOv3 on the cloud; inputting the unmanned aerial vehicle vehicle data set into S_YOLOv3 and YOLOv3 respectively, and performing model migration training, so that S_YOLOv3 meets the predetermined speed requirements, and YOLOv3 meets the predetermined accuracy requirements; the collected unmanned aerial vehicle traffic video is preliminarily filtered through the two-stage filter deployed on the unmanned aerial vehicle side, and redundant frames are discarded; the remaining frames are input into the lightweight vehicle detection model S_YOLOv3 to further filter the target frames; the target frames are input into the high-precision target detection model YOLOv3 deployed on the cloud, and high-precision detection is performed to obtain the final target frame. This invention detects unmanned aerial vehicle targets through an edge-cloud collaborative mechanism, solves the problems of time extension and target frame loss caused by excessive data transmission, and improves the target detection speed and accuracy.

[0005] CN115825088A discloses a method for fault detection of substation based on drone dual-light, wherein the system includes a fault detection data terminal, a drone dispatch terminal, a drone control terminal and a drone body; the drone dispatch terminal is used to generate drone dispatch information according to the fault detection plan preset by the inspector for the substation; the drone control terminal is used to control a corresponding number of drone bodies to perform fault detection on the substation according to the drone dispatch information; the drone body is equipped with a visible light camera and an infrared thermal imaging camera to obtain visible light image information and thermal imaging image information at the detection area in the substation: the fault detection data terminal is used to receive and analyze visible light image information and thermal imaging image information from the drone body. This invention has the effect of improving the efficiency and accuracy of fault detection.

[0006] CN202010045032.9 discloses a substation switch temperature detection method based on UAV dual-light image fusion, including: for one group of detection objects, the switch in the visible light image is identified, and the local image of the area where each identified switch device is located is marked with an identification frame; the edge contour of the switch device in each local image is identified, and the position information of each switch contact point in the visible light image is obtained according to the characteristics of the edge contour of the switch device; the position information of each switch contact point in the infrared image is calculated by combining the mapping relationship between the visible light image and the infrared image in the same group of detection objects; the temperature information of each switch contact point in the infrared image is calculated. This invention improves the speed of switch identification and temperature detection in infrared photos, can obtain temperature detection results faster, handle and respond to possible faults or abnormal situations more promptly, and improve the overall detection efficiency of substation switch temperature.

[0007] CN202210977736.9 discloses a method for detecting defects of insulators based on dual-light fusion of unmanned aerial vehicles, including: obtaining an insulator video to be detected and parsing it frame by frame to obtain a picture; wherein the insulator video includes a visible light video and an infrared light video, and after parsing frame by frame, a visible light picture and an infrared light picture are obtained respectively; performing insulator target detection on the visible light picture to determine the marking information of the insulator including the position coordinates, length, width and tilt angle; affine the insulator marking information in the visible light picture to the infrared light picture to determine the position coordinates, length, width and tilt angle of the insulator in the infrared light picture; performing insulator defect detection on the visible light picture; performing infrared thermal analysis of the insulator on the infrared light picture to obtain the temperature information of the insulator, and determining whether the insulator is overheated and causes defects by the relative temperature difference judgment method; judging whether the insulator has defects according to a preset program. Rapid detection of insulator defects is achieved, and inspection efficiency is improved. Summary of the invention

[0008] The present invention provides a ground vehicle target detection algorithm for unmanned aerial vehicle dual-light aerial images, including a dual-light fusion backbone feature extraction network, a dual-light fusion enhanced feature extraction network with weighted bidirectional features, and a dual-light fusion target detection head network; First, the UAV pod is used to collect real-time visible light and thermal imaging data of the ground observation area; data preprocessing and data enhancement are performed based on the collected video information; ground targets are detected; and the established dual-light target detection algorithm is trained; the backbone feature extraction network extracts features from visible light images and thermal imaging images respectively, and the feature maps extracted at different depths of the backbone network are fused respectively to obtain dual-light fusion feature maps at different depths, so that the ground vehicle target is more prominent in the feature map; the enhanced feature extraction network uses the dual-light fusion feature maps obtained by the backbone network at three different depths to further extract and process features, so as to achieve more accurate and robust detection of targets using dual-light fusion information; the feature map output by the enhanced feature extraction network of dual-light fusion with weighted bidirectional features is input into the target detection head, and the detection head predicts the category probability and bounding box coordinates of the target based on the dual-light fusion feature map with richer features output by the enhanced feature extraction network of dual-light fusion with weighted bidirectional features, so as to achieve accurate marking and framing of the category and position of the ground vehicle; finally, the video with the target category and detection frame is returned to the ground computer.

[0009] This application provides an algorithm for detecting ground vehicle targets in dual-light aerial images of unmanned aerial vehicles, including: 1. Obtain sample data: Use a drone dual-light pod to shoot visible light and thermal imaging ground vehicle videos, collect a large amount of training sample data, convert the videos into each frame of image, and finally label each sample data to complete the production of training data sets; 2. Construct a dual-light fusion backbone feature extraction network: Build a dual-light fusion backbone feature extraction network to perform preliminary feature extraction on the received video images and fuse the visible light and thermal imaging feature maps; The dual-light fusion backbone feature extraction network includes two parallel backbone feature extraction networks, which process two spectral images respectively, extract features through the C3 module composed of three convolutional layers and multiple Bottleneck modules based on YOLOv5, and perform feature fusion at a specific layer; Among them, after the above convolution operation, the ReLU activation function is used to increase the nonlinearity of the network, and the feature map is downsampled through the maximum pooling layer to reduce the number of parameters while retaining important features and improving the robustness of the model; 3. Construct an enhanced feature extraction network with dual-light fusion and weighted bidirectional features: Build an enhanced feature extraction network with weighted bidirectional features to perform enhanced feature extraction on the received fused dual-light feature map; The network first adjusts the resolution of the feature map through upsampling to match the shallow feature map. Then, the deep and shallow feature maps are effectively fused using the splicing operation of the Bidirectional Feature Pyramid Network (BiFPN). In the splicing process, a weighting mechanism is introduced to assign different weights to feature maps according to their importance, thereby enhancing the representation of key features and suppressing the interference of irrelevant features; Among them, the fused feature maps are further refined and extracted through multiple C3 modules to extract more robust and discriminative features, providing strong support for subsequent target detection; 4. Build a dual-light fusion target detection head network: Build a dual-light fusion target detection head network, which consists of multiple convolutional layers, upsampling layers (Upsample), feature fusion modules such as BiFPN_Concat2 and BiFPN_Concat3, and a Detect module; Among them, the feature map output by the enhanced feature extraction network with weighted bidirectional features is further processed and scaled through the convolution layer and the upsampling layer; Among them, the Detect module performs target detection on the scale-adjusted feature map and outputs information such as the target category, location, and confidence level; 5. Perform image preprocessing and data enhancement on the sample data to obtain a training set and a test set; Among them, the preprocessing uses operations such as 2x and 4x scaling, 90, 180, and 270-degree rotation, translation, and cropping of the image; This application uses the following for image enhancement: (1) Rotation and flipping: By randomly rotating and flipping the image (horizontally or vertically), the model can learn objects at different angles and directions. (2) Scaling and cropping: These operations can change the scale and size of the image, which helps the model recognize objects of different sizes. For example, the image can be scaled to different sizes or randomly cropped; (3) Image smoothing: Reduce image noise and make the image smoother by applying a smoothing filter; (4) Affine transformation: By combining linear transformation with translation, the model can adapt to complex changes such as image rotation, scaling, and shearing; (5) Translation: By moving the image in a straight line along a specific direction, the model can learn the stability of features under position changes; 6. Parameters of the dual-light fusion backbone feature extraction network and the dual-light fusion enhanced feature extraction network with weighted bidirectional features trained by the training set, wherein the test set is used to test the performance of the backbone feature extraction network model and the dual-light fusion enhanced feature extraction network model with weighted bidirectional features on unseen data; 7. The visible light image and thermal imaging image collected by the drone are simultaneously input into the dual-light fusion backbone feature extraction network, and deep feature fusion and enhancement processing is performed through the enhanced feature extraction network with weighted bidirectional features. Finally, after detection by the dual-light fusion target detection head network, the vehicle detection result is output, thereby achieving accurate identification and detection of vehicles under complex environmental conditions.

[0010] The beneficial effects of adopting the above technical solution are: This application uses the dual-light pod of an unmanned aerial vehicle to collect real-time data from the observation area, performs data preprocessing and data enhancement on the data, and improves the efficiency of the algorithm. In addition, the data is first input into the dual-light fusion trunk feature extraction network, and after the dual-light fusion feature map is extracted, it is input into the dual-light fusion enhanced feature extraction network with weighted bidirectional features to further extract small-scale target image features, thereby improving the accuracy of the algorithm. In addition, the above-mentioned dual-light fusion trunk feature extraction network and the dual-light fusion enhanced feature extraction network with weighted bidirectional features use a lightweight residual structure, which reduces network parameters and optimizes network performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This application provides a schematic diagram of the process of detecting ground vehicle targets using dual-light imaging of drones; Figure 2 This application provides a schematic diagram of the dual-light fusion backbone feature extraction network for unmanned aerial vehicles; Figure 3This application provides a schematic diagram of an enhanced feature extraction network for UAV dual-light fusion with weighted bidirectional features; Figure 4 This application provides a schematic diagram of a UAV dual-light fusion bidirectional feature pyramid network (BiFPN); Figure 5 This is the vehicle target detection result diagram of the algorithm proposed in this application on the image of the UAV dual-light image dataset VEDAI dataset. DETAILED DESCRIPTION

[0012] In order to facilitate a more straightforward further understanding and use of the present application, the present application is clearly and completely described below in conjunction with the accompanying drawings and embodiments to increase the understanding of the present application by technical personnel and users. This embodiment is only a partial illustrative embodiment of the present application. If the beneficiary individuals or units obtain other embodiments of the present application without creative labor, they are all within the scope of the rights of the present application and are protected.

[0013] Figure 1 A schematic diagram of the dual-light image ground vehicle target detection process is provided for this application.

[0014] Step S201 first uses the UAV dual-light pod to collect the ground dual-stream real-time images and saves them to the server, and pre-processes the images on the server side, wherein the specific steps include: (1) Histogram equalization: By redistributing the pixel values ​​of the image, the histogram of the image becomes more uniform, thereby enhancing the contrast and details of the image; (2) Sharpening: The sharpening algorithm can enhance the edges and details of the image; (3) Image enhancement filter: By applying a specific enhancement filter, edge enhancement filter, the edges or details of the image can be highlighted, improving the clarity and detail visibility of the image; (4) Scale the image by 2 or 4 times; (5) Rotate the image by 90 degrees, 180 degrees, 270 degrees, translate, crop, etc.

[0015] Step S202 is to construct a dual-light fusion backbone feature extraction network; Figure 2 is a schematic diagram of a dual-light fusion backbone feature extraction network provided by the present application; The network uses a dual-stream input structure to process visible light and thermal imaging images respectively, and performs initial feature extraction and slice rearrangement through the Focus layer. Subsequently, both input streams pass through multiple layers of Conv and C3 modules to gradually deepen and extract their respective features; Among them, in the key layers such as the 5th and 15th layers, the 7th and 17th layers, and the 10th and 20th layers, the network designed the Add feature fusion module to achieve the effective integration of visible light and thermal imaging feature maps;

[0016] Step S203 is to construct a dual-light fusion enhanced feature extraction network with weighted bidirectional features; Figure 3 is a schematic diagram of an enhanced feature extraction network with dual-light fusion and weighted bidirectional features provided by the present application; Among them, the fused feature map realizes deep refinement and scale restoration of the feature map through the C3 module and the upsampling module; Among them, the Bidirectional Feature Pyramid Network (BiFPN) network is used to achieve weighted bidirectional feature fusion, dynamically adjust and fuse feature maps of different scales, and improve the richness and accuracy of feature representation; Among them, the enhanced feature extraction network further optimizes the scale and receptive field of the feature map through fine convolution and upsampling operations.

[0017] Step S204 is to construct a dual-light target detection head network; The head network uses preset anchor boxes and category information to perform bounding box regression and category classification on each potential target on the feature map. Among them, the head network outputs the detection results including the target bounding box coordinates, category labels and confidence scores, realizing efficient and accurate detection of targets in complex environments.

[0018] Step S205 is the video image with the target category and detection frame finally obtained by the system.

[0019] The algorithm flow of the drone dual-light imaging ground vehicle target detection provided by the present application includes: using the drone dual-light pod to collect visible light and thermal imaging video images of the ground, inputting the collected dual-stream video images into the dual-light fusion trunk feature extraction network respectively, inputting the dual-light fusion feature map output by the dual-light fusion trunk feature extraction network into the dual-light fusion enhanced feature extraction network with weighted bidirectional features for enhanced feature extraction, inputting the enhanced feature extraction feature map of the dual-light fusion into the dual-light fusion target detection head network, and outputting the category and detection frame of the ground target in the video. Finally, the video image with the target category and detection frame is transmitted back to the ground computer.

[0020] Combine the following Figure 1 The specific implementation described in this application is only used to explain the related application, but not to limit the application. It should also be noted that, for the convenience of description, only the parts related to the application are shown in the drawings.

[0021] The server is equipped with two NVIDIA V100 graphics cards. The drone dual-light image ground vehicle target detection algorithm processes the video image in three stages. For the dual-light backbone feature extraction network, this application uses a feature extraction network composed of Focus, Conv, C3 and Add modules, such as Figure 2 As shown, it is a schematic diagram of the dual-light backbone feature extraction network of the present application. The network consists of a visible light backbone feature extraction network (301), a thermal imaging backbone feature extraction network (302), and an Add module (303). The two backbone feature extraction networks have powerful feature extraction capabilities and deep network structures, and are often used in the field of computer vision. They can better extract image features, and then the visible light feature map and the thermal imaging feature map are added and fused pixel by pixel through the Add module. The Add module fusion formula is shown in formula (1), where Y is the output tensor, x_1 and x_2 are the visible light and infrared feature map tensors respectively, and the sizes of the three fused feature maps are: 160*160, 80*80, and 40*40 respectively; Y = x _1 +x_2 (1) In actual tests, the vehicle area in the dual-light fusion feature map output by the dual-light backbone feature extraction network is clearly visible under the fusion of the thermal imaging feature map.

[0022] In the enhanced feature extraction network of dual-light fusion with weighted bidirectional features, it is necessary to perform enhanced feature extraction on dual-light feature maps of different scales. The enhanced feature extraction network of dual-light fusion with weighted bidirectional features of the present application is as follows: Figure 3 As shown, it is composed of an enhanced feature extraction module (501) and a weighted bidirectional feature pyramid network (BIFPN) fusion module (502), and through enhanced feature extraction and feature fusion, enhanced extraction and effective fusion of visible light and infrared light image features are achieved. Figure 4 The figure shows the weighted bidirectional feature pyramid network (BIFPN) structure used in the enhanced feature extraction network. The enhanced feature extraction network structure consists of three parts: bidirectional feature extraction (601), weighted feature fusion (602), and skip connection (603).

[0023] Module 602 realizes efficient feature extraction and fusion by constructing a learnable weighted fusion module. In the BiFPN_Add2 and BiFPN_Add3 classes, two and three input features are processed respectively, and they are weighted by trainable weights. The BiFPN_Add2 and BiFPN_Add3 classes are composed of 1x1 convolution and SiLU activation function; the weights are first normalized to ensure stability, then the weighted features are added, and then the number of channels is adjusted through the 1x1 convolution layer, and finally the SiLU activation function is applied to introduce nonlinearity. This structure can effectively capture multi-scale information and improve the performance of the model; its calculation expression is shown in formula (2), where w1′, w′2, and w′3 represent the weight results obtained by BiFPN_Add1, BiFPN_Add2, and BiFPN_Add3 respectively; P represents the input feature layer.

[0024] Module 602 performs weighted summation on two and three input features in the BiFPN_Add2 and BiFPN_Add3 classes, respectively, and the weights are trained by inheriting the nn.Parameter function. The fusion process first normalizes the weights to ensure that their sum is 1, then performs weighted operations on the input features and performs feature mapping through a 1x1 convolutional layer. Finally, the SiLU activation function is used to enhance nonlinear characteristics to improve model performance.

[0025] Module 603 performs cross-layer weighted summation on the 25th and 22nd layers, the 29th and 21st layers, the 32nd, 28th and 22nd layers of the dual-light fusion backbone feature extraction network through BiFPN_Add2, and performs cross-layer weighted summation on the 32nd, 28th and 22nd layers through BiFPN_Add3, thereby realizing cross-layer weighted summation of feature maps.

[0026] In order to prove the target detection effect of the dual-light fusion model of this application and verify the role of specific improvement measures, this application uses the proposed algorithm and other mainstream target detection algorithms to conduct comparative experiments in the VEDAI data set. The overall results of the experiment are shown in Table 1: Table 1: VEDAI dataset test results The experimental results are shown in Table 1, where P represents the prediction precision, R represents the prediction recall, and mAP@0.5 represents the average precision when the IoU threshold is 0.5. By comparing the detection result indicators of different models in Table 1, it can be seen that the proposed algorithm for detecting ground vehicles in dual-light aerial images of unmanned aerial vehicles has a comprehensive detection accuracy of 4.8% for 8 types of ground targets in the dataset, and mAP@0.5 is improved by 0.101 compared with the recall rate of the original YOLOv5 model based on a single light source. The proposed algorithm for detecting ground vehicles in dual-light aerial images of unmanned aerial vehicles has better indicators than other classic target detection algorithms. At the same time, Figure 5 It can be seen that in scene one, when the target appearance features are obvious, the UAV dual-light aerial image ground vehicle target detection algorithm proposed in this application can detect the target. In scene two, when the target is small, the UAV dual-light aerial image ground vehicle target detection algorithm proposed in this application can detect relatively small targets thanks to the efficient feature fusion and multi-scale information transmission capabilities of the weighted bidirectional feature pyramid network (BIFPN) structure. In scene three, since some cars in the image are obscured by shadows, the appearance information of one of the cars is almost completely missing, and YOLO-AddBI can detect the obscured target with the supplement of infrared image features. In summary, when the target in the aerial image is small and partially obscured, the classic target detection model is prone to missed detection. The UAV dual-light aerial image ground vehicle target detection algorithm proposed in this application can effectively detect the target under the same problem.

[0027] In order to verify the effectiveness of the weighted bidirectional feature pyramid network (BIFPN) structure used in this application in solving the one-way information propagation and feature information loss in the feature fusion process, ablation experiments were performed on each fusion module one by one. Some images of the VEDAI dataset were used as experimental objects. The experimental results are shown in Table 2. Table 2 Experimental results of ablation of different levels of BIFPN in the UAV dual-light image ground vehicle target detection algorithm

[0028] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) and the technical solutions formed.

Claims

1. An algorithm for detecting ground vehicle targets from dual-light aerial images of unmanned aerial vehicles, characterized in that , specifically including the following steps: S1: Obtain visible light and thermal imaging videos as sample data; S2: Preprocess and enhance the sample data, and divide it into training set, test set and validation set; S3: construct dual-light fusion backbone feature extraction network; S4: Construct an enhanced feature extraction network with dual-light fusion and weighted bidirectional features; S5: Build a dual-light fusion target detection head network; S6: Use the training set to train network parameters and the test set to evaluate model performance; S7: Input visible light and thermal imaging images to the backbone network, and after enhanced network processing, the detection head outputs the vehicle detection results.

2. The ground vehicle target detection algorithm for UAV dual-light aerial images according to claim 1 is characterized in that: The step S2 comprises: 1) Preprocess the video image; 2) Perform data augmentation on dataset images; 3) Randomly enlarge the dataset images; 4) Divide the preprocessed data set into training set, validation set and test set according to the set ratio.

3. The ground vehicle target detection algorithm for UAV dual-light aerial images according to claim 1 is characterized in that: Wherein said step S3 comprises: 1) Use two identical convolutional neural networks to perform preliminary feature extraction on the input visible light and infrared light images respectively; 2) Outputting feature maps of different scales corresponding to visible light and thermal imaging at different depths of the two convolutional neural networks; 3) The feature maps of different scales corresponding to the above-mentioned output visible light and thermal imaging are fused using the Add module.

4. The ground vehicle target detection algorithm for UAV dual-light aerial images according to claim 1 is characterized in that: Wherein said step S4 comprises: 1) First, the fusion feature map output by the dual-light fusion backbone feature extraction network is upsampled and integrated by splicing the shallow features corresponding to the backbone to enrich the features after dual-light fusion; 2) Enhance feature extraction of the integrated features and further learn the target features, thereby improving the network's ability to perceive the target; 3) Further feature extraction and upsampling are performed on the features obtained by enhanced feature extraction, and a multi-scale feature pyramid is constructed to enable the network to perceive target information of different scales at the same time.

5. The ground vehicle target detection algorithm for UAV dual-light aerial images according to claim 1 is characterized in that: The data enhancement in step 2) includes: 1) Rotation and flipping: By randomly rotating and flipping the image (horizontally or vertically), the model can learn objects at different angles and directions; 2) Scaling and cropping: These operations can change the scale and size of the image, which helps the model recognize objects of different sizes; 3) Image smoothing: Reduce image noise by applying a smoothing filter and make the image smoother; 4) Affine transformation: By combining linear transformation with translation, the model can adapt to complex changes such as rotation, scaling, and shearing of the image; 5) Translation: By moving the image in a straight line along a specific direction, the model can learn the stability of features under position changes.

6. The ground vehicle target detection algorithm for UAV dual-light aerial images according to claim 1 is characterized in that: The dual-light fusion backbone feature extraction network of step S3 includes: 1) Perform preliminary feature extraction on the input visible light and infrared images respectively, through two parallel streams, namely visible light flow and thermal imaging light flow. Each stream starts with a Focus module, slices the image and extracts preliminary features through a convolutional layer, and outputs a feature map with 64 channels; 2) In the parallel streams of visible light stream and thermal imaging stream, convolutional layers (Conv) and C3 modules are used alternately for downsampling and feature extraction, gradually deepening the network and reducing the resolution of feature maps, while increasing the number of channels to capture deeper features; after each specific layer (P3 / 8, P4 / 16, P5 / 32), the Add operation is used to fuse the feature maps of the two streams to enhance the model's ability to integrate different spectral information; 3) At the end of the visible light flow and thermal imaging flow, the spatial pyramid pooling (SPP) module is introduced to process the feature map in a multi-scale manner to enhance the model's detection ability for targets of different scales; finally, the fused features are further refined and enhanced through multiple C3 modules to provide strong feature support for subsequent enhanced feature extraction and target detection.

7. The ground vehicle target detection algorithm for UAV dual-light aerial images according to claim 1 is characterized in that: The step S4 of fusing dual light with enhanced feature extraction network with weighted bidirectional features includes: 1) First, the feature map of the two streams fused by the dual-light fusion backbone feature extraction network is upsampled; specifically, the fused feature map is upsampled to match the resolution of the shallow feature map, and the upsampled feature map is effectively integrated with the corresponding shallow feature map through the custom weighted bidirectional feature pyramid network concatenation 2 module BiFPN_Concat2 and weighted bidirectional feature pyramid network concatenation 3 module BiFPN_Concat3 concatenation operations; 2) Further refine and extract the feature maps concatenated by the weighted bidirectional feature pyramid network concatenation module BiFPN_Concat, and use multiple C3 modules to deeply mine the fused multi-scale feature maps to extract more robust and discriminative features; Construct a multi-scale feature pyramid to enhance target detection capabilities; after feature refinement, a multi-scale feature pyramid is constructed through a series of convolutional layers and upsampling operations; this structure enables the model to capture target information of different scales at the same time, thereby achieving accurate detection of ground vehicle targets.

Images