Lightweight aerial image detection method and system

Abstract

The application provides a light aerial image detection method and system, which comprises the following steps: acquiring a video frame shot by a UAV in real time; determining whether a pedestrian or vehicle target exists in the video frame by using a trained aerial image detection model; the construction process of the aerial image detection model comprises the following steps: building a backbone feature extraction network for feature extraction; building a deep feature map cross-path fusion network; splicing the built backbone feature extraction network and deep feature map cross-path fusion network and building a detection head to obtain a preset aerial image detection model; the training process of the aerial image detection model comprises the following steps: training the preset aerial image detection model by using a training set to obtain a finally converged aerial image detection model. By constructing a light aerial image detection model, the application realizes higher detection precision and lower calculation cost in aerial image detection, reduces the loss of hardware devices, and realizes light aerial image detection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image detection, and particularly relates to a light aerial image detection method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] With the continuous application of unmanned aerial vehicles in modern life, aerial photography technology has been widely used in various fields such as transportation construction, geological exploration, ecological research, and military operations. Ground vehicle target detection, as an important part of the intelligent transportation system, can provide a basis for the system's external environment perception. Through aerial photography technology, ground vehicle targets can be located and tracked, which can more clearly convey and reflect ground traffic information and traffic conditions, and provide help for building a mature intelligent transportation system.

[0004] Unmanned aerial vehicle aerial image targets are often small and dense, and the detection accuracy of existing detection algorithms is low. Moreover, the network model of the detection algorithm is too large, which will adversely affect the algorithm deployment of the unmanned aerial vehicle. All of these have brought great challenges to the unmanned aerial vehicle aerial image target detection task. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application provides a light aerial image detection method and system, which solves the problems of low detection accuracy and high computational cost of existing aerial image detection algorithms, and realizes significant improvement in detection accuracy and lightweight aerial image target detection.

[0006] To achieve the above purpose, one or more embodiments of the present application provide the following technical solutions:

[0007] The first aspect of the present application provides a light aerial image detection method, comprising:

[0008] real-time acquisition of unmanned aerial vehicle shooting video frames;

[0009] using the trained aerial image detection model to determine whether there are pedestrian or vehicle targets in the video frames;

[0010] The construction process of the aerial image detection model comprises:

[0011] building a backbone feature extraction network for feature extraction;

[0012] building a deep feature map cross-path fusion network;

[0013] splicing the built backbone feature extraction network and deep feature map cross-path fusion network and building a detection head to obtain a preset aerial image detection model;

[0014] The training process of the aerial image detection model comprises:

[0015] The preset aerial image detection model is trained by using the training set to obtain a finally converged aerial image detection model.

[0016] The second aspect of the present application provides a light aerial image detection system, comprising:

[0017] The video frame acquisition module is configured to acquire real-time UAV video frames.

[0018] The aerial image detection model construction module is configured to build a backbone feature extraction network for feature extraction.

[0019] A deep feature map cross-path fusion network is built.

[0020] The built backbone feature extraction network and deep feature map cross-path fusion network are spliced and a detection head is built to obtain a preset aerial image detection model.

[0021] The aerial image detection model training module is configured to train the preset aerial image detection model by using the training set to obtain a finally converged aerial image detection model.

[0022] The target judgment module is configured to judge whether there are pedestrian or vehicle targets in the video frames by using the trained aerial image detection model.

[0023] The third aspect of the present application provides a computer readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a light aerial image detection method according to the first aspect of the present application.

[0024] The fourth aspect of the present application provides an electronic device comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of a light aerial image detection method according to the first aspect of the present application.

[0025] The above one or more technical solutions have the following beneficial effects:

[0026] The present application builds a light aerial image target detection model, i.e. a backbone feature extraction network comprising a VB network and a C3SFN network, and a DFM-CPFN (Deep Feature Map Cross Path Fusion Network) as a multi-scale feature fusion network, which realizes higher detection accuracy and lower calculation and parameter amount of aerial image target detection, solves the problems of existing aerial image detection algorithms such as complexity, low precision and high cost, realizes light-weight high-precision aerial image detection, and has wide applicability.

[0027] Advantages of additional aspects of the present application will be set forth in part in the description which follows, and in part will be obvious from the description, or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0028] The accompanying drawings, which form a part of the specification, are included to provide a further understanding of the application and are incorporated herein by reference. The illustrations are shown schematically and are not intended to limit the application.

[0029] Figure 1 A flow chart of the light aerial image detection method in embodiment one of the present application;

[0030] Figure 2 (a) and Figure 2 (b) are respectively the VB network and C3SFN network structure diagrams of the backbone feature extraction network in embodiment one of the present application;

[0031] Figure 3 A network structure diagram of the deep feature map cross path fusion network DFM-CPFN (Deep Feature Map Cross Path Fusion Network) in embodiment one of the present application;

[0032] Figure 4 A comparison result diagram of the aerial image detection algorithm (LAI-YOLOv5s) in embodiment one of the present application and the YOLO v5m algorithm, the YOLOv5l algorithm and the YOLO v3 algorithm in the mAP@0.5 index;

[0033] Figure 5 A comparison result diagram of the aerial image detection algorithm (LAI-YOLOv5s) in embodiment one of the present application and the YOLO v5m algorithm, the YOLOv5l algorithm and the YOLO v3 algorithm in the mAP@0.5:0.95 index. DETAILED DESCRIPTION

[0034] Embodiment one

[0035] As Figure 1 shown, the embodiment discloses a light aerial image detection method, comprising:

[0036] Obtaining a plurality of unmanned aerial vehicle aerial images with pedestrians or vehicles, and pre-processing the aerial images;

[0037] Labeling the pedestrians or vehicles and other targets in the aerial images, taking a plurality of labeled pictures as a training sample set, and training a preset aerial image detection model;

[0038] Real-time acquisition of unmanned aerial vehicle shooting video frames, using the trained aerial image detection model to judge whether there are pedestrians or vehicles in the video frames.

[0039] In the embodiment, first, 6471 unmanned aerial vehicle aerial images with pedestrians or vehicles are acquired, which are all color images stored in JPG format. The color unmanned aerial vehicle aerial images are preprocessed to improve the quality of the unmanned aerial vehicle aerial images and the image reading speed through traditional image processing technology and Mosaic data enhancement technology, specifically including:

[0040] First, the existing image processing technology is used to process the unmanned aerial vehicle aerial images, including but not limited to random flipping, splicing, scaling, etc., to improve the data set quality and improve the model generalization ability. Second, the Mosaic data enhancement technology is used to process the unmanned aerial vehicle aerial images to improve the reading speed of the aerial images. Finally, the size of the unmanned aerial vehicle aerial images is normalized to the same size, and in the embodiment, the size of the aerial images is normalized to 640x640, which is convenient for subsequent unified processing.

[0041] Then, the labeling tool is used to label the pedestrians or vehicles in the preprocessed unmanned aerial vehicle aerial images, i.e., the labels such as 'pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', and'motor' (corresponding to "pedestrian", "people", "bicycle", "car", "van", "truck", "tricycle", "awning-tricycle", "bus", and "motor", respectively) are labeled on the upper left of the pedestrians or vehicles in the images, and a square frame is used to frame each corresponding target. For all pedestrians or vehicles, five parameters are determined, including the category parameter, the target center coordinate parameter x, y, and the target width and height parameter w, h, wherein the target center coordinate and the target width and height are determined relative to the overall picture. The label file of the unmanned aerial vehicle aerial image data is converted to txt format for storage. A number of labeled unmanned aerial vehicle aerial images are used as a training sample set to train the preset unmanned aerial vehicle aerial image detection model.

[0042] As another embodiment, 80% of the aerial images are selected to constitute a training set, 10% of the aerial images are selected to constitute a validation set, and 10% of the aerial images are selected to constitute a test set. The labeling tool is used to label the pedestrians or vehicles in the training set of unmanned aerial vehicle aerial images, and the training set is used to train the preset unmanned aerial vehicle aerial image detection model.

[0043] The construction process of the preset unmanned aerial vehicle aerial image detection model specifically includes:

[0044] A backbone network for feature extraction is built.

[0045] A deep feature map cross-path fusion network (DFM-CPFN) and a detection head are built.

[0046] The built backbone network and the deep feature map cross-path fusion network are spliced, and the spliced network is trained by using a training sample set to obtain a finally converged network model.

[0047] Existing aerial image detection algorithms usually adopt a YOLO v5 network structure, and the YOLO v5 network structure includes a backbone feature extraction network and a multi-scale feature fusion network.

[0048] In the YOLO v5 network structure, the backbone feature extraction network is a huge network structure. It adopts a C3 module combining the CSP idea and the residual network idea, and uses a residual block stacking mode of "3, 6, 9, 3" combination in the module. Each C3 module is connected by a down-sampling connection with a convolution kernel size of 3 and a step size of 2, gradually reducing the image size and increasing the receptive field.

[0049] In this embodiment, the C3 module is optimized and improved. First, a VB network is designed by combining VoVNet and residual structure Bottleneck, which can make the network narrower, the transmission of features and gradients more effective, and can improve the degradation problem of the neural network and the gradient disappearance problem caused by deepening the network. Then, a C3SFN network is designed by combining the idea of ShuffleNetV2, which greatly reduces the calculation cost of the original C3 network and makes the entire network model more lightweight.

[0050] The unmanned aerial vehicle aerial image labeled with pedestrian or vehicle targets is input into the backbone feature extraction network. In this embodiment, the backbone feature extraction network adopts 3 VB networks and 1 C3SFN network, and a CBS convolution module with a convolution kernel size of 1 and a step size of 2 is used to connect the two networks.

[0051] As shown in Figure 2 The VB network starts with 4 CBS convolution modules, the output of each convolution module is spliced after the last module, and the Bottleneck module is reserved as a separate path for splicing. After all the splicing operations are completed, a convolution module is used to adjust the channel number of the output.

[0052] As shown in Figure 2(b) as shown, the C3SFN network is replaced by a SFN module instead of the Bottleneck module in the original C3 network, the SFN module is a Bottleneck-like module, which has two branches, one branch contains two ordinary convolutions with a kernel size of 1 and a depth separable convolution with a kernel size of 3, the other branch is the input, the two branches are fused by a concatenation operation, and finally a Channel Shuffle operation is used to enable information communication between the two branches. Through the above backbone feature extraction network, the number of channels of the feature map is changed to obtain feature maps with different channels.

[0053] In addition, the present embodiment abandons the original residual block stacking mode of "3, 6, 9, 3" of the YOLO v5 network structure, and adopts a stacking mode of "3, 3, 3, 6"; "3, 3, 3" corresponds to three VB modules, that is, the Bottleneck modules in the three VBs are stacked 3 respectively; "6" corresponds to a C3SFN module, that is, the SFN module in the C3SFN is stacked 6. Through the improved stacking mode, the calculation cost of the model can be effectively reduced under the premise of ensuring that the final recognition accuracy loss is very small.

[0054] In the YOLO v5 network structure, the multi-scale feature extraction network Neck learns the fusion idea of PANet, and adds a bottom-up feature fusion path compared with FPN; moreover, the reverse optimization method (i.e. the loss function is back propagated) adopted by YOLO v5 mainly includes center coordinate prediction, confidence prediction and classification prediction. Among them, the center coordinate prediction uses the Sigmoid function to limit the output value to 0-1, which is beneficial to the optimization of the output parameters, and the confidence prediction and classification prediction use binary cross entropy for back propagation prediction.

[0055] In the present embodiment, in order to further improve the feature extraction effect and reduce the amount of calculation, as shown in Figure 3 As shown, based on the multi-scale feature fusion network of PANet, a deep feature map cross path fusion network DFM-CPFN (Deep Feature Map Cross Path Fusion Network) is designed. First, the path and concatenation operation after the medium size detection head are removed and replaced by two upsampling operations and two C3 modules to introduce a larger size feature map for detecting smaller targets; secondly, a concatenation operation is added after the two upsampling operations, which is cross path fusion with the main network. It can enhance the fusion ability of deep and shallow information, enrich the semantic information of bottom features, and improve the detection effect. At the same time, in order to make the network model more lightweight, the C3SFN network is also applied to the DFM-CPFN.

[0056] After that, the scheme described in this embodiment still uses the reverse optimization method of YOLO v5 for center coordinate prediction, confidence prediction and classification prediction.

[0057] The labeled image is input into the backbone feature extraction network, and four feature maps F1, F2, F3 and F4 of different channels are output. The four feature maps are input into the DFM-CPFN respectively, and different scale feature maps D1, D2 and D3 are obtained. The most suitable prediction box is screened out through non-maximum suppression (NMS) and IoU intersection over union. DFM-CPFN is a feature fusion idea rather than a fixed module network structure. The specific input or fusion method of the four maps in this example can be seen in the attached Figure 3 ;

[0058] In this embodiment, each grid in the feature map D1, D2 and D3 contains 3 prediction boxes, and the channel number of each prediction box is 15. The first channel is the result of confidence prediction, which is a probability value. The second to fifth channels are the results of center coordinate prediction, which output the prediction box relative to the entire feature map. The sixth to fifteenth channels are the results of classification prediction, which are also probability values.

[0059] After the backbone feature extraction network and the DFM-CPFN are built, the built backbone feature extraction network is spliced with the built DFM-CPFN. That is, the first, second, third and fourth output feature maps of the backbone network are used as input for multi-DFM-CPFN, and the output feature maps are different. The convolution operation of the DFM-CPFN outputs three detection feature maps, including a 40x40 scale (D2) feature map for detecting large targets, an 80x80 (D1) scale feature map for detecting medium targets, and a 160x160 (D3) scale feature map for detecting small targets.

[0060] The above-mentioned preset aerial image detection model is trained based on the training sample set. Under the premise of ensuring that overfitting and underfitting do not occur, the aerial image detection model is fully trained until the loss function no longer converges, the optimal weight file is retained, and the training is completed.

[0061] Actually, the above-mentioned aerial image detection model of this embodiment includes an input layer, an intermediate layer and an output layer. The images in the training sample set are input into the input layer of the detection model. The intermediate layer includes a backbone network and a DFM-CPFN. The VB and C3SFN network structures are stacked in the backbone network. The DFM-CPFN contains cross-path feature fusion methods from top to bottom and from bottom to top. Feature extraction is performed through the intermediate layer, and finally three feature maps of different scales are output in the output layer to determine the probability of whether the input training image contains pedestrian or vehicle targets.

[0062] Finally, the real-time acquisition of the unmanned aerial vehicle aerial video frame is obtained, the aerial image detection model after training is used for judgment, whether the target such as pedestrian or vehicle exists in the video frame, and the probability of the target is output.

[0063] In order to fully prove the effectiveness of the above-mentioned detection method of the embodiment, the YOLO v3 algorithm and the YOLO v5 algorithm are trained in the same experimental environment, and the corresponding mAP@0.5 (the value of the IoU threshold is greater than 0.5, the higher the value, the greater the threshold, and the lower the precision), mAP@0.5:0.95 (the average mAP of multiple thresholds from 0.5 to 0.95 with a step of 0.05), the amount of calculation and the amount of parameters are obtained, which are used to prove the effectiveness of the improved scheme proposed in the embodiment.

[0064] In order to prove the superiority of the aerial image detection algorithm (LAI-YOLOv5s) described in the embodiment in comprehensive performance, comparative experiments are carried out on the detection effect and calculation cost of the aerial image detection algorithm and YOLOv5m, YOLOv51 and YOLOv3. The comparative results of the aerial image detection algorithm (LAI-YOLOv5s) and the YOLO v5m algorithm, the YOLO v5l algorithm and the YOLOv3 algorithm on the mAP@0.5 index are as shown in Table 1, the comparative results on the mAP@0.5:0.95 index are as shown in Table 2, and the comparative results on the amount of calculation and the amount of parameters are as shown in Table 3. Figure 4 Figure 5

[0065] The comparative results of the aerial image detection algorithm described in the embodiment and the detection effect and calculation cost of YOLOv5m, YOLOv51 and YOLOv3 are as shown in Table 1. Among them, the amount of parameters and the amount of calculation of the aerial image detection algorithm described in the embodiment are only 30.1% and 60.4% of YOLOv5m, the mAP@0.5 index is increased by 4.5%, and the mAP@0.5:0.95 index is increased by 2.2%. Compared with YOLOv5l, the amount of parameters and the amount of calculation of the aerial image detection algorithm described in the embodiment are only 13.6% and 26.9% of YOLOv5l, the mAP@0.5 index is increased by 2.7%, and the mAP@0.5:0.95 index is increased by 0.8%. Compared with YOLOv3, the amount of parameters and the amount of calculation of the aerial image detection algorithm described in the embodiment are only 10.2% and 18.7% of YOLOv3, the mAP@0.5 index is increased by 1.4%, and the mAP@0.5:0.95 index is increased by 0.5%. The experiment proves that the aerial image detection algorithm proposed in the embodiment has better technical effect, which is not only very light in calculation cost, but also has higher detection precision.

[0066] Table 1 ​​

[0067]

[0068] In summary, the embodiment provides a lightweight aerial image detection method, which solves the problems of low detection accuracy and high calculation cost of existing detection algorithms, and realizes lightweight high-precision aerial image detection.

[0069] Embodiment two

[0070] The embodiment discloses a lightweight aerial image detection system, comprising:

[0071] The video frame acquisition module is configured to acquire real-time unmanned aerial vehicle shooting video frames;

[0072] The aerial image detection model construction module is configured to build a backbone feature extraction network for feature extraction;

[0073] A deep feature map cross-path fusion network is built;

[0074] The built backbone feature extraction network and deep feature map cross-path fusion network are spliced and a detection head is built to obtain a preset aerial image detection model;

[0075] The aerial image detection model training module is configured to train the preset aerial image detection model using a training set to obtain a finally converged aerial image detection model;

[0076] The target judgment module is configured to judge whether there are pedestrian or vehicle targets in the video frames using the trained aerial image detection model.

[0077] Embodiment three

[0078] The purpose of the embodiment is to provide a computer readable storage medium.

[0079] The computer readable storage medium has a computer program stored thereon, and the program is executed by a processor to implement the steps in the lightweight aerial image detection method according to Embodiment 1 of the present disclosure.

[0080] Embodiment four

[0081] The purpose of the embodiment is to provide an electronic device.

[0082] The electronic device includes a memory, a processor, and a program stored on the memory and executable on the processor, and the processor executes the program to implement the steps in the lightweight aerial image detection method according to Embodiment 1 of the present disclosure.

[0083] The steps involved in the apparatuses of the above embodiments two, three and four correspond to the method of embodiment one, and the specific implementation can refer to the relevant description of embodiment one. The term "computer readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; it should also be understood as including any medium capable of storing, encoding or carrying the instruction set for execution by the processor and causing the processor to perform any of the methods in the present application.

[0084] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively manufactured into each integrated circuit module, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0085] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the protection scope of the present application, and those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A lightweight aerial image detection method, characterized in that, include: Real-time acquisition of video frames captured by drones; The trained aerial image detection model is used to determine whether there are pedestrian or vehicle targets in the video frame; The construction process of the aerial image detection model includes: Construct a backbone feature extraction network for feature extraction: In the YOLO v5 network architecture, the C3 module is optimized and improved. The backbone feature extraction network includes a VB network and a C3SFN network. The VB network is obtained by fusing VoVNet and the residual structure Bottleneck. The VB network starts with four CBS convolutional modules, and the output of each convolutional module is concatenated after the last module. The Bottleneck module is retained as a separate concatenation path. After all concatenation operations are completed, one more convolutional module is used to adjust the number of output channels. The C3SFN network is obtained by replacing the Bottleneck module in the original C3 network with an SFN module. The SFN module is a Bottleneck-like module with two branches. One branch contains two ordinary convolutions with a kernel size of 1 and a depthwise separable convolution with a kernel size of 3. The other branch is the input. The two branches are fused together by a concatenation operation, and finally, a Channel is used. Shuffle operation enables communication between the two branches; the backbone feature extraction network uses three VB networks and one C3SFN network, which are connected by a CBS convolutional module with a kernel of 1 and a stride of 2. Building a deep feature map cross-path fusion network (DFM-CPFN): In the YOLO v5 network architecture, a deep feature map cross-path fusion network is designed based on the PANet multi-scale feature fusion network. This deep feature map cross-path fusion network employs a dual-path feature fusion approach, combining top-down and bottom-up methods. The labeled image is first input into the backbone feature extraction network, which outputs four feature maps F1, F2, F3, and F4 from different channels. F1, F2, and F3 are the outputs of the first, second, and third VB modules in the backbone feature extraction network, respectively, while F4 is the final output. The final output F4 is then input into the multi-scale feature fusion network DFM-CPFN, ultimately generating three different-sized detection heads D1, D2, and D3 for detecting targets of varying sizes. During the multi-scale feature fusion process, the shallow features F1 and the intermediate features F2 and F3 from the backbone feature extraction network all participate. Finally, the model selects the most suitable prediction box using non-maximum suppression and IoU (Intersection over Union). In the first bottom-up feature fusion path, the feature map size is progressively enlarged by two upsampling operations. After the first upsampling, its output is fused with the intermediate feature F3 of the backbone feature extraction network. Then, after the second upsampling, its output is fused with the intermediate feature F2 of the backbone feature extraction network. The fused features are used in the forward pass. The output of the first feature fusion path serves as the first detection head D1 of the model and is also the initial input of the second feature fusion path. In the second top-down fusion path, a downsampling convolution is first used to reduce the feature size, and then it is fused with the output of the CBS convolution module of the first fusion path. After passing through the C3 module, it becomes the second detection head D2 of the model. D2 is smaller in size than D1 and has a larger receptive field, making it suitable for detecting larger targets. Subsequently, the feature map corresponding to D2 is subjected to two upsampling operations and deeper convolution extraction. After the first upsampling, its output is fused with the intermediate feature F2 of the backbone feature extraction network. After the second upsampling, it is fused with the shallow feature F1 of the backbone feature extraction network to obtain richer information such as target location and coordinates. The final deep cross-path fusion feature is used as the third detection head D3 of the model. Its size is larger than D1 and has a smaller receptive field, making it suitable for detecting small targets. The backbone feature extraction network and the deep feature map cross-path fusion network are spliced ​​together to build a detection head, thus obtaining the preset aerial image detection model. The training process of the aerial image detection model includes: The preset aerial image detection model is trained using the training set to obtain the final converged aerial image detection model.

2. The lightweight aerial image detection method as described in claim 1, characterized in that, The process of constructing the training set includes: acquiring several aerial images containing pedestrians or vehicles, and preprocessing the aerial images; The pedestrians and vehicles in the preprocessed aerial images are labeled, and a portion of the labeled aerial images are used as the training set.

3. The lightweight aerial image detection method as described in claim 2, characterized in that, The preprocessing of aerial images includes: First, the aerial images are randomly flipped, stitched, and scaled. Secondly, Mosaic data augmentation technology is used to process aerial images; Finally, the aerial images were normalized to the same size.

4. A lightweight aerial image detection system, characterized in that: include: The video frame acquisition module is configured to acquire video frames captured by the drone in real time. The aerial image detection model building module is configured to: build a backbone feature extraction network for feature extraction; In the YOLO v5 network architecture, the C3 module is optimized and improved. The backbone feature extraction network includes a VB network and a C3SFN network. The VB network is obtained by fusing VoVNet and the residual structure Bottleneck. The VB network starts with four CBS convolutional modules, and the output of each convolutional module is concatenated after the last module. The Bottleneck module is retained as a separate concatenation path. After all concatenation operations are completed, one more convolutional module is used to adjust the number of output channels. The C3SFN network is obtained by replacing the Bottleneck module in the original C3 network with an SFN module. The SFN module is a Bottleneck-like module with two branches. One branch contains two ordinary convolutions with a kernel size of 1 and a depthwise separable convolution with a kernel size of 3. The other branch is the input. The two branches are fused together by a concatenation operation, and finally, a Channel is used. Shuffle operation enables communication between the two branches; the backbone feature extraction network uses three VB networks and one C3SFN network, which are connected by a CBS convolutional module with a kernel of 1 and a stride of 2. Building a deep feature map cross-path fusion network (DFM-CPFN): In the YOLO v5 network architecture, a deep feature map cross-path fusion network is designed based on the PANet multi-scale feature fusion network. The deep feature map cross-path fusion network adopts a dual-path feature fusion method of top-down and bottom-up. The labeled image is first input into the backbone feature extraction network, which outputs four feature maps F1, F2, F3, and F4 from different channels. F1, F2, and F3 are the outputs of the first, second, and third VB modules in the backbone feature extraction network, respectively, while F4 is the final output. The final output F4 is then input into the multi-scale feature fusion network DFM-CPFN, ultimately generating three different sized detection heads D1, D2, and D3 for detecting targets of varying sizes. During the multi-scale feature fusion process, the shallow features F1 and the intermediate features F2 and F3 from the backbone feature extraction network are all involved. Finally, the model selects the most suitable prediction box using non-maximum suppression and IoU (Intersection over Union). In the first bottom-up feature fusion path, the feature map size is progressively enlarged by two upsampling operations. After the first upsampling, its output is fused with the intermediate feature F3 of the backbone feature extraction network. Then, after the second upsampling, its output is fused with the intermediate feature F2 of the backbone feature extraction network. The fused features are used in the forward pass. The output of the first feature fusion path serves as the first detection head D1 of the model and is also the initial input of the second feature fusion path. In the second top-down fusion path, a downsampling convolution is first used to reduce the feature size, and then it is fused with the output of the CBS convolution module of the first fusion path. After passing through the C3 module, it becomes the second detection head D2 of the model. D2 is smaller in size than D1 and has a larger receptive field, making it suitable for detecting larger targets. Subsequently, the feature map corresponding to D2 is subjected to two upsampling operations and deeper convolution extraction. After the first upsampling, its output is fused with the intermediate feature F2 of the backbone feature extraction network. After the second upsampling, it is fused with the shallow feature F1 of the backbone feature extraction network to obtain richer information such as target location and coordinates. The final deep cross-path fusion feature is used as the third detection head D3 of the model. Its size is larger than D1 and has a smaller receptive field, making it suitable for detecting small targets. The backbone feature extraction network and the deep feature map cross-path fusion network are spliced ​​together to build a detection head, thus obtaining the preset aerial image detection model. The aerial image detection model training module is configured to: train the preset aerial image detection model using the training set to obtain the final converged aerial image detection model; The target determination module is configured to use a trained aerial image detection model to determine whether there are pedestrian or vehicle targets in the video frame.

5. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps of a lightweight aerial image detection method as described in any one of claims 1-3.

6. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the lightweight aerial image detection method as described in any one of claims 1-3.

Images