A method for detecting ground objects using drones

Abstract

This invention provides a method for detecting ground objects using drones, belonging to the field of ground object recognition in drone aerial photography. The method comprises: acquiring aerial images captured by a visible light camera; using an information entropy-based ambient light brightness perception algorithm to sense the ambient light brightness level based on the acquired aerial images; and selecting a corresponding deep learning network to detect ground pedestrians and vehicles based on the perceived ambient light brightness level. This method enables rapid and highly robust detection of ground vehicles and pedestrians in visible light camera aerial images under complex ambient lighting conditions.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) aerial photography ground object recognition, and in particular to a UAV ground object detection method. Background Art

[0002] In recent years, drones have been widely used in public safety applications such as disaster relief, security monitoring, and emergency response. Drones equipped with visible light cameras can achieve high-precision reproduction of scene objects, offering rich detail, intuitive visualization, and ease of understanding. However, visible light cameras are also highly susceptible to ambient light and drone motion, resulting in uneven brightness, overly bright scenes, or dark scenes. This can significantly impact the design and application of object detection algorithms.

[0003] To solve the above problems, the industry currently adopts the following solutions:

[0004] 1) Improving the quality of camera input light by installing a filter in front of the visible light camera lens. The main problem with this method is that the optical information removed by the filter can easily affect the quality of the camera output image, thereby affecting the visual experience.

[0005] 2) By using digital image processing methods, image enhancement and noise reduction are performed to improve the quality of the collected images. The main problem with the above methods is that most of the algorithms currently designed are not robust and are prone to significant performance degradation once the ambient light changes.

[0006] 3) By using advanced neural network technology and combining it with massive amounts of training data, a high detection rate algorithm can be designed. The above method can significantly improve the accuracy of detection and recognition, but its main drawback is that accumulating massive amounts of data and labeling the targets is time-consuming and labor-intensive, making it difficult to implement in practice.

[0007] Visible light camera imaging environment perception technology primarily refers to specialized modeling and simulation techniques that use image light field distribution estimation to analyze the direction and intensity of ambient lighting in a scene, thereby guiding the design and calculation of subsequent algorithms. Imaging environment perception technology can include two-dimensional light field environment perception and three-dimensional light field environment perception. While the computational accuracy of two-dimensional light field environment perception is not as high as that of three-dimensional light field environment perception, it offers faster computation speed, lower computational resource consumption, and better real-time performance. Furthermore, when used in conjunction with technologies such as neural networks, the perception and classification accuracy can meet the requirements of most engineering applications, making it particularly suitable for application in the field of drone-based object detection and recognition technology. Summary of the Invention

[0008] The present invention provides a method for detecting ground objects using drones, which can achieve rapid and robust detection of ground vehicles and pedestrians in aerial images taken by visible light cameras under complex ambient lighting conditions. The technical solution is as follows:

[0009] An embodiment of the present invention provides a method for detecting ground objects using a drone, comprising:

[0010] Acquire aerial images from a visible light camera;

[0011] According to the acquired aerial images, the ambient light brightness level is perceived using an ambient light brightness perception algorithm based on information entropy;

[0012] According to the perceived ambient light brightness level, the corresponding deep learning network is selected for ground pedestrian and vehicle detection.

[0013] Furthermore, the step of sensing the ambient light brightness level based on the acquired aerial image by using an ambient light brightness perception algorithm based on information entropy includes:

[0014] Convert the real-time acquired environmental image from RGB color space to LAB color space;

[0015] Calculate the mean loss contrast normalization index MSCN of the L component, A component and B component in the LAB color space L 、MSCN A and MSCN B ;

[0016] Calculating MSCN L 、MSCN A and MSCN B The brightness information entropy of the three components H L 、H A and H B ;

[0017] A support vector machine classifier is used to evaluate the ambient light brightness level of the aerial image, wherein the input of the support vector machine classifier is H L 、H A and H B , the output is the ambient light brightness level;

[0018] The ambient light brightness levels include: relatively strong brightness ambient light, relatively weak brightness ambient light and extremely weak brightness ambient light.

[0019] Furthermore, selecting a corresponding deep learning network to detect pedestrians and vehicles on the ground according to the perceived ambient light brightness level includes:

[0020] For three different levels of ambient light, three improved Yolov3-tiny lightweight networks were constructed for ground vehicles and pedestrians, for a total of six improved networks, to detect ground vehicles and people under different sizes, shapes, and brightness levels of ambient light.

[0021] Under strong ambient light conditions, the Yolov3-tiny improved network Yolov3-tiny-VDB1 is constructed for ground vehicle detection;

[0022] Under weak ambient light conditions, the Yolov3-tiny improved network Yolov3-tiny-VDB2 is constructed for ground vehicle detection;

[0023] Under extremely weak ambient light conditions, the Yolov3-tiny improved network Yolov3-tiny-VDB3 is constructed for ground vehicle detection;

[0024] Under strong ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB1 was constructed to detect ground personnel.

[0025] Under weak ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB2 is constructed to detect ground personnel;

[0026] Under extremely weak ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB3 is constructed for ground personnel detection.

[0027] Furthermore, the Yolov3-tiny improved network Yolov3-tiny-VDB1 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a third convolutional layer, a first route layer, a fourth convolutional layer, a second route layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, a third route layer, a seventh convolutional layer, a fourth route layer, a third pooling layer, an eighth convolutional layer, a ninth convolutional layer, a fifth route layer, a tenth convolutional layer, a sixth route layer, a fourth pooling layer, an eleventh convolutional layer, a twelfth convolutional layer, a first shortcut layer, a thirteenth convolutional layer, a first Yolo prediction layer, a fourteenth convolutional layer, an upsampling layer, a second shortcut layer, a fifteenth convolutional layer, a third shortcut layer, a fourth shortcut layer, a sixteenth convolutional layer and a second Yolo prediction layer; wherein,

[0028] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the third convolutional layer, the first route layer, the fourth convolutional layer, the second route layer, the second pooling layer, the fifth convolutional layer, the sixth convolutional layer, the third route layer, the seventh convolutional layer, the fourth route layer, the third pooling layer, the eighth convolutional layer, the ninth convolutional layer, the fifth route layer, the tenth convolutional layer, the sixth route layer, the fourth pooling layer, the eleventh convolutional layer, the twelfth convolutional layer, the first shortcut layer, the thirteenth convolutional layer, the first Yolopredictio n are connected in sequence, the second convolutional layer is also connected to the first route layer and the second route layer respectively, the fifth convolutional layer is also connected to the third route layer and the fourth route layer respectively, the eighth convolutional layer is also connected to the fifth route layer and the sixth route layer respectively, the eleventh convolutional layer is also connected to the first shortcut layer, the fourteenth convolutional layer, the upsampling layer, the second shortcut layer, the fifteenth convolutional layer, the third shortcut layer, the fourth shortcut layer, the sixteenth convolutional layer and the second Yolo prediction are connected in sequence, the upsampling layer is also connected to the third shortcut layer, the sixth route layer is also connected to the fourth shortcut layer, and the third pooling layer is also connected to the second shortcut layer;

[0029] The input resolution of the input layer is 512×512.

[0030] Furthermore, the Yolov3-tiny improved network Yolov3-tiny-VDB2 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a third pooling layer, a sixth convolutional layer, a seventh convolutional layer, a third route layer, an eighth convolutional layer, a fourth route layer, a ninth convolutional layer, a fifth route layer, a fourth pooling layer, a tenth convolutional layer, an eleventh convolutional layer, an SPP layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a first Yolo prediction layer, a sixth route layer, a fifteenth convolutional layer, an upsampling layer, a seventh route layer, a sixteenth convolutional layer, a seventeenth convolutional layer and a second Yolo prediction layer; wherein,

[0031] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the third pooling layer, the sixth convolutional layer, the seventh convolutional layer, the third route layer, the eighth convolutional layer, the fourth route layer, the ninth convolutional layer, the fifth route layer, the fourth pooling layer, the tenth convolutional layer, the eleventh convolutional layer, the SPP layer, the twelfth convolutional layer, the thirteenth convolutional layer, the fourteenth convolutional layer, and the first Yolo prediction are connected in sequence, the third convolutional layer is also connected to the first route layer and the second route layer respectively, the sixth convolutional layer is also connected to the third route layer, the fourth route layer, and the fifth route layer respectively, the twelfth convolutional layer is also connected to the sixth route layer, the sixth route layer, the fifteenth convolutional layer, the upsampling layer, the seventh route layer, the sixteenth convolutional layer, the seventeenth convolutional layer, and the second Yolo prediction are connected in sequence, and the ninth convolutional layer is also connected to the seventh route layer;

[0032] The input resolution of the input layer is 512×512.

[0033] Furthermore, the Yolov3-tiny improved network Yolov3-tiny-VDB3 includes: an input layer, a first convolutional layer, a first pooling layer, a dense block, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a fifth convolutional layer, an SPP layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a first Yolo prediction, a first route layer, a ninth convolutional layer, an upsampling layer, a second route layer, a tenth convolutional layer, an eleventh convolutional layer, and a second Yolo prediction; wherein,

[0034] The input layer, the first convolution layer, the first pooling layer, the Denseblock, the second convolution layer, the second pooling layer, the third convolution layer, the third pooling layer, the fourth convolution layer, the fourth pooling layer, the fifth convolution layer, the SPP layer, the sixth convolution layer, the seventh convolution layer, the eighth convolution layer, and the first Yolo prediction are connected in sequence, the sixth convolution layer is connected to the first route layer and the seventh convolution layer, the ninth convolution layer, the upsampling layer, the second route layer, the tenth convolution layer, the eleventh convolution layer, and the second Yolo prediction are connected in sequence, and the fourth convolution layer is also connected to the second route layer;

[0035] The input resolution of the input layer is 512×512.

[0036] Furthermore, the Yolov3-tiny improved network Yolov3-tiny-PDB1 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a third pooling layer, a sixth convolutional layer, a seventh convolutional layer, a third route layer, an eighth convolutional layer, a fourth route layer, a fourth pooling layer, a ninth convolutional layer, a tenth convolutional layer, a first shortcut layer, an eleventh convolutional layer, a first Yolo prediction, a fifth route layer, a twelfth convolutional layer, a first upsampling layer, a second shortcut layer, a thirteenth convolutional layer, a third shortcut layer, a fourth shortcut layer, a fourteenth convolutional layer, a second Yolo prediction, a fifteenth convolutional layer, a second upsampling layer, a sixth route layer, a sixteenth convolutional layer and a third Yolo prediction; wherein,

[0037] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the third pooling layer, the sixth convolutional layer, the seventh convolutional layer, the third route layer, the eighth convolutional layer, the fourth route layer, the fourth pooling layer, the ninth convolutional layer, the tenth convolutional layer, the first shortcut layer, the fifth route layer, the eleventh convolutional layer, the first upsampling layer, the second shortcut layer, the twelfth convolutional layer, the third shortcut layer, the fourth shortcut layer, the thirteenth convolutional layer, the second Yolo layer The third convolutional layer is further connected to the first routing layer and the second routing layer respectively, the second routing layer is further connected to the sixth routing layer, the sixth convolutional layer is further connected to the third routing layer and the fourth routing layer respectively, the third routing layer is further connected to the second shortcut layer, the fourth routing layer is further connected to the fourth shortcut layer, the first upsampling layer is further connected to the third shortcut layer, the ninth convolutional layer is further connected to the first shortcut layer, the first shortcut is further connected to the first Yolo prediction through the fourteenth convolutional layer, and the fourth routing layer is further connected to the third Yolo prediction through the fifteenth convolutional layer, the second upsampling layer, the sixth routing layer, and the sixteenth convolutional layer in sequence;

[0038] The input resolution of the input layer is 544×544.

[0039] Furthermore, the Yolov3-tiny improved network Yolov3-tiny-PDB2 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a sixth convolutional layer, a third route layer, a seventh convolutional layer, a fourth route layer, an eighth convolutional layer, a fifth route layer, a ninth convolutional layer, a sixth route layer, a third pooling layer, a tenth convolutional layer, an SPP layer, an eleventh convolutional layer, a fourth pooling layer, a twelfth convolutional layer, a fifth pooling layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a first Yolo prediction layer, a sixteenth convolutional layer, a first upsampling layer, a seventh route layer, a seventeenth convolutional layer, an eighteenth convolutional layer, a second Yolo prediction layer, a nineteenth convolutional layer, a second upsampling layer, an eighth route layer, a twentieth convolutional layer and a third Yolo prediction layer; wherein,

[0040] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the sixth convolutional layer, the third route layer, the seventh convolutional layer, the fourth route layer, the eighth convolutional layer, the fifth route layer, the ninth convolutional layer, the sixth route layer, the third pooling layer, the tenth convolutional layer, the SPP layer, the eleventh convolutional layer, the fourth pooling layer, the twelfth convolutional layer, the fifth pooling layer, the thirteenth convolutional layer, the fourteenth convolutional layer, the fifteenth convolutional layer, and the first Yolo prediction are connected in sequence, and the third convolutional layer is also connected to the first route, the second route, the third route, the fourth route, the fifth route, and the sixth route respectively. The thirteenth convolutional layer is also connected to the second Yolo prediction through the sixteenth convolutional layer, the first upsampling layer, the seventh route layer, the seventeenth convolutional layer, and the eighteenth convolutional layer. The eleventh convolutional layer is connected to the seventh routing layer and the nineteenth convolutional layer respectively, and the nineteenth convolutional layer is connected to the third Yoloprediction through the second upsampling layer, the eighth routing layer, and the twentieth convolutional layer;

[0041] The input resolution of the input layer is 544×544.

[0042] Furthermore, the Yolov3-tiny improved network Yolov3-tiny-PDB3 includes: an input layer, a first convolutional layer, a first pooling layer, a dense block, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first route layer, a sixth convolutional layer, a second route layer, a first SPP layer, a seventh convolutional layer, a third pooling layer, a second SPP layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, a first Yolo prediction layer, an eleventh convolutional layer, a first upsampling layer, a third route layer, a twelfth convolutional layer, a thirteenth convolutional layer, a second Yolo prediction layer, a fourteenth convolutional layer, a second upsampling layer, a fourth route layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a seventeenth convolutional layer and a third Yolo prediction layer; wherein,

[0043] The input layer, the first convolutional layer, the first pooling layer, the Denseblock, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the first route layer, the sixth convolutional layer, the second route layer, the first SPP layer, the seventh convolutional layer, the third pooling layer, the second SPP layer, the eighth convolutional layer, the ninth convolutional layer, the tenth convolutional layer, and the first Yoloprediction are connected in sequence, the third convolutional layer is also connected to the second route layer, the fourth convolutional layer is also connected to the first route layer, the second route layer is also connected to the seventeenth convolutional layer, the seventh convolutional layer is also connected to the third route layer and the fourteenth convolutional layer respectively, the eighth convolutional layer is also connected to the second Yoloprediction through the eleventh convolutional layer, the first upsampling layer, the third route layer, the twelfth convolutional layer, the thirteenth convolutional layer and the second Yoloprediction. prediction; the fourteenth convolutional layer is connected to the third Yoloprediction through the second upsampling layer, the fourth routing layer, the fifteenth convolutional layer, and the sixteenth convolutional layer in sequence, and the seventeenth convolutional layer is also connected to the fourth routing layer;

[0044] The input resolution of the input layer is 544×544.

[0045] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0046] In this embodiment of the present invention, aerial images captured by a visible light camera are acquired. Based on the acquired aerial images, an ambient light brightness perception algorithm based on information entropy is used to sense the ambient light brightness level. Based on the perceived ambient light brightness level, a corresponding deep learning network is selected to detect pedestrians and vehicles on the ground. This classification of the ambient light brightness level in the images effectively reduces the amount of training data required for deep learning network training for single-type images. This approach offers the advantages of simple and efficient network training and minimal training data requirements. This allows for rapid and robust detection of vehicles and pedestrians on the ground in aerial images captured by visible light cameras under complex ambient lighting conditions, with a high detection rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 A schematic diagram of the flow of a method for detecting ground objects using a drone according to an embodiment of the present invention;

[0049] Figure 2 A detailed flowchart of the drone ground object detection method provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of the structure of the original Yolov3-tiny network provided in an embodiment of the present invention;

[0051] Figure 4 A schematic diagram of the structure of the Yolov3-tiny improved network Yolov3-tiny-VDB1 provided in an embodiment of the present invention;

[0052] Figure 5 A schematic diagram of the structure of the Yolov3-tiny improved network Yolov3-tiny-VDB2 provided in an embodiment of the present invention;

[0053] Figure 6 A schematic diagram of the structure of the Yolov3-tiny improved network Yolov3-tiny-VDB3 provided in an embodiment of the present invention;

[0054] Figure 7 A schematic diagram of the structure of the Yolov3-tiny improved network Yolov3-tiny-PDB1 provided in an embodiment of the present invention;

[0055] Figure 8A schematic diagram of the structure of the Yolov3-tiny improved network Yolov3-tiny-PDB2 provided in an embodiment of the present invention;

[0056] Figure 9 A schematic diagram of the structure of the Yolov3-tiny improved network Yolov3-tiny-PDB3 provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0058] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a method for detecting ground objects using a drone, comprising:

[0059] S101, acquiring an aerial image taken by a visible light camera;

[0060] S102, based on the acquired aerial image, using an ambient light brightness perception algorithm based on information entropy to perceive the ambient light brightness level, which may specifically include the following steps:

[0061] A1, converts the real-time acquired environment image from RGB color space to LAB color space;

[0062] A2, calculate the mean subtracted and contrast normalized (MSCN) index of the L component, A component, and B component in the LAB color space L 、MSCN A and MSCN B ;

[0063] In this embodiment, MSCN X Indicator express, The calculation method is as follows:

[0064]

[0065]

[0066]

[0067] Among them, I X (i, j) represents the image intensity of the X component at position (i, j), X∈{L,A,B}, i=1,2,…,M, j=1,2,…,N, M, N represent the length and width of the image respectively; C is a constant, generally C=1; w={w k,l|k=-K,…,K;l=-L,…,L} is a two-dimensional circularly symmetric Gaussian weighting function, K and L represent the template size values of the discrete Gaussian function in the horizontal and vertical directions, respectively, preferably K=L=2; μ X (i, j) represents the weighted grayscale value of the image; σ X (i,j) represents the variance value of the image.

[0068] A3, calculate MSCN L 、MSCN A and MSCN B The brightness information entropy of the three components H L 、H A and H B ;

[0069] In this embodiment, the brightness information entropy component H X Expressed as:

[0070]

[0071] Among them, m represents the grayscale value of [0,255], X∈{L,A,B}, P X (m) represents the proportion of pixels with gray value m in all pixels of the image. The image here is taken as the image obtained by MSCN. X The calculated image.

[0072] A4, using a support vector machine (SVM) classifier to evaluate the ambient light brightness level of the aerial image, wherein the input of the support vector machine classifier is H L 、H A and H B , the output is the ambient light brightness level;

[0073] The ambient light brightness level is divided into three levels, including: relatively strong brightness ambient light, relatively weak brightness ambient light and extremely weak brightness ambient light.

[0074] In this embodiment, the ambient light brightness level of the aerial image is determined according to the result of SVM classification, wherein the number of training data levels of SVM should be greater than 1000.

[0075] S103: Select a corresponding deep learning network to detect pedestrians and vehicles on the ground based on the perceived ambient light brightness level.

[0076] like Figure 3As shown, the original Yolov3-tiny network prediction branches only support large and medium sizes of 13×13 and 26×26, making them less suitable for detecting small-sized targets. Furthermore, while deep learning networks offer better computational performance than traditional neural networks, they require massive amounts of data for training in order to increase their generalization capabilities, which is time-consuming and labor-intensive in practical applications. Therefore, to address these issues to a certain extent, this embodiment implements the following improved Yolov3-tiny network design for ambient light with varying brightness levels and different ground objects. By classifying image types, this effectively reduces the amount of training data required for deep learning network training for single-type images, significantly improving the practicality of this embodiment.

[0077] In this embodiment, for three different levels of ambient light, three types of improved Yolov3-tiny lightweight networks are constructed for ground vehicles and pedestrians, for a total of six types of improved networks, to detect ground vehicles and people under different sizes, shapes, and different levels of ambient light.

[0078] 1) Under strong ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-VDB1 is constructed to detect ground vehicles;

[0079] In this embodiment, the Yolov3-tiny improved network Yolov3-tiny-VDB1 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a third convolutional layer, a first route layer, a fourth convolutional layer, a second route layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, a third route layer, a seventh convolutional layer, a fourth route layer, a third pooling layer, an eighth convolutional layer, a ninth convolutional layer, a fifth route layer, a tenth convolutional layer, a sixth route layer, a fourth pooling layer, an eleventh convolutional layer, a twelfth convolutional layer, a first shortcut layer, a thirteenth convolutional layer, a first Yolo prediction layer, a fourteenth convolutional layer, an upsampling layer, a second shortcut layer, a fifteenth convolutional layer, a third shortcut layer, a fourth shortcut layer, a sixteenth convolutional layer and a second Yolo prediction layer; wherein,

[0080] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the third convolutional layer, the first route layer, the fourth convolutional layer, the second route layer, the second pooling layer, the fifth convolutional layer, the sixth convolutional layer, the third route layer, the seventh convolutional layer, the fourth route layer, the third pooling layer, the eighth convolutional layer, the ninth convolutional layer, the fifth route layer, the tenth convolutional layer, the sixth route layer, the fourth pooling layer, the eleventh convolutional layer, the twelfth convolutional layer, the first shortcut layer, the thirteenth convolutional layer, the first Yolopredictio n are connected in sequence, the second convolutional layer is also connected to the first route layer and the second route layer respectively, the fifth convolutional layer is also connected to the third route layer and the fourth route layer respectively, the eighth convolutional layer is also connected to the fifth route layer and the sixth route layer respectively, the eleventh convolutional layer is also connected to the first shortcut layer, the fourteenth convolutional layer, the upsampling layer, the second shortcut layer, the fifteenth convolutional layer, the third shortcut layer, the fourth shortcut layer, the sixteenth convolutional layer and the second Yolo prediction are connected in sequence, the upsampling layer is also connected to the third shortcut layer, the sixth route layer is also connected to the fourth shortcut layer, and the third pooling layer is also connected to the second shortcut layer;

[0081] The input resolution of the input layer is 512×512.

[0082] In this embodiment, the relatively strong brightness ambient light is defined as the ambient light brightness>1500lx, or the brightness information entropy three components H L 、H A , HB satisfies H L >4.8、2.8>H A >1.5、2.5>H B >1.0.

[0083] In this embodiment, first, the input resolution of the Yolov3-tiny improved network Yolov3-tiny-VDB1 is set to 512×512. Under strong ambient light conditions, most aerial vehicles are small and medium-sized targets and are very sensitive to position information deviation. Increasing the network input resolution will enable the detected targets to have more accurate position information and richer detail information; secondly, the two prediction branches of the original Yolov3-tiny network are removed, and a bidirectional feature pyramid network (BiFPN) structure (such as Figure 4The inner part of the dotted box in the middle) fuses the features from the deep and shallow layers at all levels, and finally sets up 32×32 and 64×64 prediction branches; Finally, in order to enhance the extraction of shallow information, 1×1 and 3×3 convolution layers are added after the second, third, and fourth convolution layers of the original Yolov3-tiny network, and the upper layer features are fused respectively; Because the feature extraction performance of 1×1 and 3×3 convolution layers is better, and they can enhance the semantic information of small targets, such as Figure 4 As shown in the figure, under good lighting conditions, the vehicle features at all levels are more obvious. The above improvements can achieve better recognition results without increasing too many network parameters.

[0084] 2) Under low-brightness ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-VDB2 is constructed to detect ground vehicles;

[0085] In this embodiment, the Yolov3-tiny improved network Yolov3-tiny-VDB2 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a third pooling layer, a sixth convolutional layer, a seventh convolutional layer, a third route layer, an eighth convolutional layer, a fourth route layer, a ninth convolutional layer, a fifth route layer, a fourth pooling layer, a tenth convolutional layer, an eleventh convolutional layer, an SPP layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a first Yolo prediction layer, a sixth route layer, a fifteenth convolutional layer, an upsampling layer, a seventh route layer, a sixteenth convolutional layer, a seventeenth convolutional layer, and a second Yolo prediction layer; wherein,

[0086] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the third pooling layer, the sixth convolutional layer, the seventh convolutional layer, the third route layer, the eighth convolutional layer, the fourth route layer, the ninth convolutional layer, the fifth route layer, the fourth pooling layer, the tenth convolutional layer, the eleventh convolutional layer, the SPP layer, the twelfth convolutional layer, the thirteenth convolutional layer, the fourteenth convolutional layer, and the first Yolo prediction are connected in sequence, the third convolutional layer is also connected to the first route layer and the second route layer respectively, the sixth convolutional layer is also connected to the third route layer, the fourth route layer, and the fifth route layer respectively, the twelfth convolutional layer is also connected to the sixth route layer, the sixth route layer, the fifteenth convolutional layer, the upsampling layer, the seventh route layer, the sixteenth convolutional layer, the seventeenth convolutional layer, and the second Yolo prediction are connected in sequence, and the ninth convolutional layer is also connected to the seventh route layer;

[0087] The input resolution of the input layer is 512×512.

[0088] In this embodiment, the relatively weak ambient light is defined as the ambient light brightness between 300lx and 1500lx, or the brightness information entropy three components H L 、H A HB meets 5.0>H L ≥3.7, 2.9>H A ≥1.7, 2.2>H B >1.3.

[0089] In this embodiment, when the ambient light conditions are weak, the features of the vehicle at all levels are relatively fuzzy. In order to strengthen the extraction of vehicle features, 1×1 and 3×3 convolutional layers are added after the third layer of the original Yolov3-tiny network, and 1×1, 3×3 and 3×3 convolutional layers are added after the fourth layer. The upper layer features are fused respectively to strengthen the information extraction of the shallow network and reduce the loss of feature information. At the same time, a spatial pyramid pooling (SPP) structure is added after the 7th convolutional layer, and a maximum pooling combination of 5×5, 9×9, and 13×13 is used to form the SPP structure. The main purpose is to extract pooled features from the feature maps using different pooling layers, and then fuse them to improve the detection accuracy. Finally, the network input resolution is also increased to 512×512, such as Figure 5 shown.

[0090] 3) Under extremely weak ambient light conditions, construct the Yolov3-tiny improved network Yolov3-tiny-VDB3 for ground vehicle detection;

[0091] In this embodiment, the Yolov3-tiny improved network Yolov3-tiny-VDB3 includes: an input layer, a first convolutional layer, a first pooling layer, a dense block, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a fifth convolutional layer, an SPP layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a first Yolo prediction, a first route layer, a ninth convolutional layer, an upsampling layer, a second route layer, a tenth convolutional layer, an eleventh convolutional layer, and a second Yolo prediction; wherein,

[0092] The input layer, the first convolution layer, the first pooling layer, the Denseblock, the second convolution layer, the second pooling layer, the third convolution layer, the third pooling layer, the fourth convolution layer, the fourth pooling layer, the fifth convolution layer, the SPP layer, the sixth convolution layer, the seventh convolution layer, the eighth convolution layer, and the first Yolo prediction are connected in sequence, the sixth convolution layer is connected to the first route layer and the seventh convolution layer, the ninth convolution layer, the upsampling layer, the second route layer, the tenth convolution layer, the eleventh convolution layer, and the second Yolo prediction are connected in sequence, and the fourth convolution layer is also connected to the second route layer;

[0093] The input resolution of the input layer is 512×512.

[0094] In this embodiment, extremely weak ambient light is defined as ambient light brightness between 50lx and 300lx, or the three components of brightness information entropy H L 、H A HB satisfies 2.8≤H L <4.8, 1.0≤H A <3.6, 1.2 <H B <2.0.

[0095] In this embodiment, the lighting conditions are weaker under extremely poor ambient light conditions, the target is not prominent, the features are not obvious, and the feature information is relatively small. Therefore, a network with a deeper structure is needed to extract vehicle features. This embodiment selects the feature extraction network of DenseNet to replace the feature extraction network of Yolov3-tiny, wherein 1 Denseblock is used, and each Denseblock is composed of 6 pairs of 1×1 and 3×3 convolution layers. The idea of Denseblock is that the input of each layer comes from the output of all previous layers. The 3×3 convolution in each Denseblock is preceded by a 1×1 convolution operation, the purpose of which is to reduce the number of input feature maps, which can not only reduce the dimension and reduce the amount of calculation, but also fuse the features of each channel. Secondly, the SPP structure is added after the DenseNet feature extraction network to better integrate and utilize feature information. Finally, the input resolution of the network is changed to 512×512, such as Figure 6 shown.

[0096] 4) Under strong ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB1 was constructed to detect ground personnel;

[0097] In this embodiment, the Yolov3-tiny improved network Yolov3-tiny-PDB1 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a third pooling layer, a sixth convolutional layer, a seventh convolutional layer, a third route layer, an eighth convolutional layer, a fourth route layer, a fourth pooling layer, a ninth convolutional layer, a tenth convolutional layer, a first shortcut layer, an eleventh convolutional layer, a first Yolo prediction layer, a fifth route layer, a twelfth convolutional layer, a first upsampling layer, a second shortcut layer, a thirteenth convolutional layer, a third shortcut layer, a fourth shortcut layer, a fourteenth convolutional layer, a second Yolo prediction layer, a fifteenth convolutional layer, a second upsampling layer, a sixth route layer, a sixteenth convolutional layer, and a third Yolo prediction layer; wherein,

[0098] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the third pooling layer, the sixth convolutional layer, the seventh convolutional layer, the third route layer, the eighth convolutional layer, the fourth route layer, the fourth pooling layer, the ninth convolutional layer, the tenth convolutional layer, the first shortcut layer, the fifth route layer, the eleventh convolutional layer, the first upsampling layer, the second shortcut layer, the twelfth convolutional layer, the third shortcut layer, the fourth shortcut layer, the thirteenth convolutional layer, the second Yolo layer The third convolutional layer is further connected to the first routing layer and the second routing layer respectively, the second routing layer is further connected to the sixth routing layer, the sixth convolutional layer is further connected to the third routing layer and the fourth routing layer respectively, the third routing layer is further connected to the second shortcut layer, the fourth routing layer is further connected to the fourth shortcut layer, the first upsampling layer is further connected to the third shortcut layer, the ninth convolutional layer is further connected to the first shortcut layer, the first shortcut is further connected to the first Yolo prediction through the fourteenth convolutional layer, and the fourth routing layer is further connected to the third Yolo prediction through the fifteenth convolutional layer, the second upsampling layer, the sixth routing layer, and the sixteenth convolutional layer in sequence;

[0099] The input resolution of the input layer is 544×544.

[0100] In this embodiment, the relatively strong brightness ambient light is defined as the ambient light brightness>1000lx, or the brightness information entropy three components H L 、H A , HB satisfies H L >5.0、2.6>H A >1.5、2.0>H B>1.0. When the ambient light conditions are good, the features of pedestrians are more obvious, but as pedestrians are small targets, feature extraction mainly relies on shallow networks. First, the input size of the image is adjusted. Since the pedestrian target is too small, the image size of 544×544 is selected as the network input to better express the longitudinal information of the pedestrian. Secondly, since most pedestrians are small targets, the receptive field size of the shallow network feature map is small, so 1×1 and 3×3 convolution layers are added after the third and fourth convolution layers of the original Yolov3-tiny network, and the upper-layer features are fused respectively to strengthen the information extraction of the shallow network, enhance the semantic information of small targets in the network, and avoid the loss of shallow feature information. Thirdly, an upsampling layer Upsample2 is added to the ninth convolution layer of the original Yolov3-tiny network, connected to the upper convolution output, and a new feature pyramid layer is added (such as Figure 7 The final feature pyramid network output is: 136×136×18, 68×68×18, 34×34×18, as shown in the figure. Figure 7 shown.

[0101] 5) Under low-brightness ambient light conditions, construct the Yolov3-tiny improved network Yolov3-tiny-PDB2 for ground personnel detection;

[0102] In this embodiment, the Yolov3-tiny improved network Yolov3-tiny-PDB2 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a sixth convolutional layer, a third route layer, a seventh convolutional layer, a fourth route layer, an eighth convolutional layer, a fifth route layer, a ninth convolutional layer, a sixth route layer, a third pooling layer, a tenth convolutional layer, an SPP layer, an eleventh convolutional layer, a fourth pooling layer, a twelfth convolutional layer, a fifth pooling layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a first Yolo prediction layer, a sixteenth convolutional layer, a first upsampling layer, a seventh route layer, a seventeenth convolutional layer, an eighteenth convolutional layer, a second Yolo prediction layer, a nineteenth convolutional layer, a second upsampling layer, an eighth route layer, a twentieth convolutional layer, and a third Yolo prediction layer; wherein,

[0103] The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the sixth convolutional layer, the third route layer, the seventh convolutional layer, the fourth route layer, the eighth convolutional layer, the fifth route layer, the ninth convolutional layer, the sixth route layer, the third pooling layer, the tenth convolutional layer, the SPP layer, the eleventh convolutional layer, the fourth pooling layer, the twelfth convolutional layer, the fifth pooling layer, the thirteenth convolutional layer, the fourteenth convolutional layer, the fifteenth convolutional layer, and the first Yolo prediction are connected in sequence, and the third convolutional layer is also connected to the first route, the second route, the third route, the fourth route, the fifth route, and the sixth route respectively. The thirteenth convolutional layer is also connected to the second Yolo prediction through the sixteenth convolutional layer, the first upsampling layer, the seventh route layer, the seventeenth convolutional layer, and the eighteenth convolutional layer. The eleventh convolutional layer is connected to the seventh routing layer and the nineteenth convolutional layer respectively, and the nineteenth convolutional layer is connected to the third Yoloprediction through the second upsampling layer, the eighth routing layer, and the twentieth convolutional layer;

[0104] The input resolution of the input layer is 544×544.

[0105] In this embodiment, the relatively weak ambient light is defined as the ambient light brightness between 600lx and 1000lx, or the brightness information entropy three components H L 、H A HB meets 5.0>H L ≥4.0, 2.6>H A ≥2.0, 1.9>H B >1.3. Pedestrian features under weaker ambient light conditions are not as clear as those under stronger ambient light conditions. At this time, while strengthening feature extraction, a prediction branch with a larger feature resolution should be added to detect small pedestrian targets under drone aerial photography. First, the network input resolution is changed to 544×544 to better express pedestrian information. Secondly, six convolutional layers of 1×1, 3×3, 1×1, 3×3, 1×1 and 3×3 are added after the third layer of the original Yolov3-tiny network, and each layer fuses features from the upper layer in turn to strengthen the information extraction of the shallow network and reduce the loss of feature information. At the same time, the SPP structure is added after the fourth convolution layer of the original Yolov3-tiny, and the maximum pooling combination of 5×5, 9×9, and 13×13 is used to form the SPP structure. Finally, a prediction branch with a feature resolution of 136×136 is added to improve the detection of small pedestrian targets in environments with weak ambient light, such as Figure 8 shown.

[0106] 6) Under extremely weak ambient light conditions, an improved Yolov3-tiny network Yolov3-tiny-PDB3 is constructed to detect ground personnel.

[0107] In this embodiment, the Yolov3-tiny improved network Yolov3-tiny-PDB3 includes: an input layer, a first convolutional layer, a first pooling layer, a dense block, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first route layer, a sixth convolutional layer, a second route layer, a first SPP layer, a seventh convolutional layer, a third pooling layer, a second SPP layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, a first Yolo prediction layer, an eleventh convolutional layer, a first upsampling layer, a third route layer, a twelfth convolutional layer, a thirteenth convolutional layer, a second Yolo prediction layer, a fourteenth convolutional layer, a second upsampling layer, a fourth route layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a seventeenth convolutional layer and a third Yolo prediction layer; wherein,

[0108] The input layer, the first convolutional layer, the first pooling layer, the Denseblock, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the first route layer, the sixth convolutional layer, the second route layer, the first SPP layer, the seventh convolutional layer, the third pooling layer, the second SPP layer, the eighth convolutional layer, the ninth convolutional layer, the tenth convolutional layer, and the first Yoloprediction are connected in sequence, the third convolutional layer is also connected to the second route layer, the fourth convolutional layer is also connected to the first route layer, the second route layer is also connected to the seventeenth convolutional layer, the seventh convolutional layer is also connected to the third route layer and the fourteenth convolutional layer respectively, the eighth convolutional layer is also connected to the second Yoloprediction through the eleventh convolutional layer, the first upsampling layer, the third route layer, the twelfth convolutional layer, the thirteenth convolutional layer and the second Yoloprediction. prediction; the fourteenth convolutional layer is connected to the third Yoloprediction through the second upsampling layer, the fourth routing layer, the fifteenth convolutional layer, and the sixteenth convolutional layer in sequence, and the seventeenth convolutional layer is also connected to the fourth routing layer;

[0109] The input resolution of the input layer is 544×544.

[0110] In this embodiment, extremely weak ambient light is defined as ambient light brightness between 150lx and 600lx, or the three components of brightness information entropy H L 、H A , HB satisfies 4.1≤H L <4.8, 2.6≤HA <3.5, 1.3 <H B <1.8. Pedestrian features under extremely weak ambient light conditions are less obvious and there are more blurred areas. Therefore, in order to extract sufficient pedestrian feature information, it is necessary not only to deepen the network depth but also to add large-resolution prediction branches. First, the input resolution is changed to 544×544 to better express pedestrian information. Secondly, a Dense Block is used to replace the second convolution layer of the original Yolov3-tiny, and 1×1, 3×3 and 1×1 convolution layers are added after the fourth layer of the original Yolov3-tiny network, and the feature information from the previous field is fused respectively to strengthen the information extraction of the shallow network and reduce the loss of feature information. Thirdly, an SPP structure is added before each prediction branch to better integrate and utilize feature information. Finally, a prediction branch with a feature resolution of 136×136 is added to improve the detection accuracy of small pedestrian targets under extremely weak ambient light conditions, such as Figure 9 shown.

[0111] The drone ground object detection method described in an embodiment of the present invention acquires aerial imagery captured by a visible light camera; based on the acquired aerial imagery, it uses an ambient light brightness perception algorithm based on information entropy to perceive the ambient light brightness level; and based on the perceived ambient light brightness level, it selects a corresponding deep learning network to detect ground pedestrians and vehicles. This classification of the ambient light brightness level in the image effectively reduces the amount of training data required for deep learning network training for single-type images. This method offers the advantages of simple and efficient network training and minimal training data requirements, thereby enabling rapid and robust detection of ground vehicles and pedestrians in visible light camera aerial imagery under complex ambient lighting conditions, with a high detection rate.

[0112] In this embodiment, a drone ground object detection system is also provided, which includes: a rotary wing or fixed wing drone equipped with a visible light camera, a wireless data transmission unit and a ground data processing unit; wherein,

[0113] The visible light camera is used to take aerial photos, obtain aerial images, and transmit the obtained aerial images to a ground data processing unit via a wireless data transmission unit;

[0114] The ground data processing unit is used to perceive the ambient light brightness level based on the received aerial images using an ambient light brightness perception algorithm based on information entropy; and is also used to select a corresponding deep learning network for ground pedestrian and vehicle detection based on the perceived ambient light brightness level.

[0115] In this embodiment, the function of the ground data processing unit may also be replaced by an airborne data processing unit.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting ground objects using an unmanned aerial vehicle, characterized in that: include: Acquire aerial images from a visible light camera; According to the acquired aerial images, the ambient light brightness level is perceived using an ambient light brightness perception algorithm based on information entropy; According to the perceived ambient light brightness level, the corresponding deep learning network is selected to detect pedestrians and vehicles on the ground; The step of sensing the ambient light brightness level based on the acquired aerial image by using an ambient light brightness perception algorithm based on information entropy includes: Convert the real-time acquired environmental image from RGB color space to LAB color space; Calculate the mean loss contrast normalization index MSCN of the L component, A component and B component in the LAB color space L 、MSCN A and MSCN B ; Calculating MSCN L 、MSCN A and MSCN B The brightness information entropy of the three components H L 、H A and H B ; A support vector machine classifier is used to evaluate the ambient light brightness level of the aerial image, wherein the input of the support vector machine classifier is H L 、H A and H B , the output is the ambient light brightness level; The ambient light brightness levels include: relatively strong ambient light, relatively weak ambient light and extremely weak ambient light; The method of selecting a corresponding deep learning network to detect pedestrians and vehicles on the ground according to the perceived ambient light brightness level includes: For three different brightness levels of ambient light, three improved Yolov3-tiny lightweight networks were constructed for ground vehicles and pedestrians, totaling six types, to detect ground vehicles and personnel under ambient light conditions of different sizes, shapes and brightness levels.

2. The method for detecting ground objects by a drone according to claim 1, wherein: The Yolov3-tiny improved network Yolov3-tiny-VDB1 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a third convolutional layer, a first route layer, a fourth convolutional layer, a second route layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, a third route layer, a seventh convolutional layer, a fourth route layer, a third pooling layer, an eighth convolutional layer, a ninth convolutional layer, a fifth route layer, a tenth convolutional layer, a sixth route layer, a fourth pooling layer, an eleventh convolutional layer, a twelfth convolutional layer, a first shortcut layer, a thirteenth convolutional layer, a first Yolo prediction layer, a fourteenth convolutional layer, an upsampling layer, a second shortcut layer, a fifteenth convolutional layer, a third shortcut layer, a fourth shortcut layer, a sixteenth convolutional layer and a second Yolo prediction layer; wherein, The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the third convolutional layer, the first route layer, the fourth convolutional layer, the second route layer, the second pooling layer, the fifth convolutional layer, the sixth convolutional layer, the third route layer, the seventh convolutional layer, the fourth route layer, the third pooling layer, the eighth convolutional layer, the ninth convolutional layer, the fifth route layer, the tenth convolutional layer, the sixth route layer, the fourth pooling layer, the eleventh convolutional layer, the twelfth convolutional layer, the first shortcut layer, the thirteenth convolutional layer, the first Yoloprediction layer The second convolutional layer is further connected to the first route layer and the second route layer respectively, the fifth convolutional layer is further connected to the third route layer and the fourth route layer respectively, the eighth convolutional layer is further connected to the fifth route layer and the sixth route layer respectively, the eleventh convolutional layer is further connected to the first shortcut layer, the fourteenth convolutional layer, the upsampling layer, the second shortcut layer, the fifteenth convolutional layer, the third shortcut layer, the fourth shortcut layer, the sixteenth convolutional layer and the second Yolo prediction layer are connected in sequence, the upsampling layer is further connected to the third shortcut layer, the sixth route layer is further connected to the fourth shortcut layer, and the third pooling layer is further connected to the second shortcut layer; The input resolution of the input layer is 512×512.

3. The method for detecting ground objects by a drone according to claim 1, wherein: The Yolov3-tiny improved network Yolov3-tiny-VDB2 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first routing layer, a fifth convolutional layer, a second routing layer, a third pooling layer, a sixth convolutional layer, a seventh convolutional layer, a third routing layer, an eighth convolutional layer, a fourth routing layer, a ninth convolutional layer, a fifth routing layer, a fourth pooling layer, a tenth convolutional layer, an eleventh convolutional layer, an SPP layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a first Yolo prediction layer, a sixth routing layer, a fifteenth convolutional layer, an upsampling layer, a seventh routing layer, a sixteenth convolutional layer, a seventeenth convolutional layer and a second Yolo prediction layer; wherein, The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first routing layer, the fifth convolutional layer, the second routing layer, the third pooling layer, the sixth convolutional layer, the seventh convolutional layer, the third routing layer, the eighth convolutional layer, the fourth routing layer, the ninth convolutional layer, the fifth routing layer, the fourth pooling layer, the tenth convolutional layer, the eleventh convolutional layer, the SPP layer, the twelfth convolutional layer, the thirteenth convolutional layer, the fourteenth convolutional layer, and the first Yolo prediction layer are connected in sequence; the third convolutional layer is also connected to the first routing layer and the second routing layer respectively; the sixth convolutional layer is also connected to the third routing layer, the fourth routing layer, and the fifth routing layer respectively; the twelfth convolutional layer is also connected to the sixth routing layer; the sixth routing layer, the fifteenth convolutional layer, the upsampling layer, the seventh routing layer, the sixteenth convolutional layer, the seventeenth convolutional layer, and the second Yolo prediction layer are connected in sequence; and the ninth convolutional layer is also connected to the seventh routing layer; The input resolution of the input layer is 512×512.

4. The method for detecting ground objects by a drone according to claim 1, wherein: The Yolov3-tiny improved network Yolov3-tiny-VDB3 includes: an input layer, a first convolutional layer, a first pooling layer, a dense block, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a fifth convolutional layer, an SPP layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a first Yolo prediction layer, a first route layer, a ninth convolutional layer, an upsampling layer, a second route layer, a tenth convolutional layer, an eleventh convolutional layer and a second Yolo prediction layer; wherein, The input layer, the first convolutional layer, the first pooling layer, the Denseblock, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the fourth convolutional layer, the fourth pooling layer, the fifth convolutional layer, the SPP layer, the sixth convolutional layer, the seventh convolutional layer, the eighth convolutional layer, and the first Yolo prediction layer are connected in sequence, the sixth convolutional layer is connected to the first route layer and the seventh convolutional layer, the ninth convolutional layer, the upsampling layer, the second route layer, the tenth convolutional layer, the eleventh convolutional layer, and the second Yolo prediction layer are connected in sequence, and the fourth convolutional layer is also connected to the second route layer; The input resolution of the input layer is 512×512.

5. The method for detecting ground objects by a drone according to claim 1, wherein: The Yolov3-tiny improved network Yolov3-tiny-PDB1 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a third pooling layer, a sixth convolutional layer, a seventh convolutional layer, a third route layer, an eighth convolutional layer, a fourth route layer, a fourth pooling layer, a ninth convolutional layer, a tenth convolutional layer, a first shortcut layer, an eleventh convolutional layer, a first Yolo prediction layer, a fifth route layer, a twelfth convolutional layer, a first upsampling layer, a second shortcut layer, a thirteenth convolutional layer, a third shortcut layer, a fourth shortcut layer, a fourteenth convolutional layer, a second Yolo prediction layer, a fifteenth convolutional layer, a second upsampling layer, a sixth route layer, a sixteenth convolutional layer and a third Yolo prediction layer; wherein, The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the third pooling layer, the sixth convolutional layer, the seventh convolutional layer, the third route layer, the eighth convolutional layer, the fourth route layer, the fourth pooling layer, the ninth convolutional layer, the tenth convolutional layer, the first shortcut layer, the fifth route layer, the eleventh convolutional layer, the first upsampling layer, the second shortcut layer, the twelfth convolutional layer, the third shortcut layer, the fourth shortcut layer, the thirteenth convolutional layer, the second Yolo layer The prediction layers are connected in sequence, the third convolutional layer is also connected to the first routing layer and the second routing layer respectively, the second routing layer is also connected to the sixth routing layer, the sixth convolutional layer is also connected to the third routing layer and the fourth routing layer respectively, the third routing layer is also connected to the second shortcut layer, the fourth routing layer is also connected to the fourth shortcut layer, the first upsampling layer is also connected to the third shortcut layer, the ninth convolutional layer is also connected to the first shortcut layer, the first shortcut layer is also connected to the first Yolo prediction layer through the fourteenth convolutional layer, and the fourth routing layer is also connected to the third Yolo prediction layer through the fifteenth convolutional layer, the second upsampling layer, the sixth routing layer, and the sixteenth convolutional layer in sequence; The input resolution of the input layer is 544×544.

6. The method for detecting ground objects by a drone according to claim 1, wherein: The Yolov3-tiny improved network Yolov3-tiny-PDB2 includes: an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a first route layer, a fifth convolutional layer, a second route layer, a sixth convolutional layer, a third route layer, a seventh convolutional layer, a fourth route layer, an eighth convolutional layer, a fifth route layer, a ninth convolutional layer, a sixth route layer, a third pooling layer, a tenth convolutional layer, an SPP layer, an eleventh convolutional layer, a fourth pooling layer, a twelfth convolutional layer, a fifth pooling layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a first Yolo prediction layer, a sixteenth convolutional layer, a first upsampling layer, a seventh route layer, a seventeenth convolutional layer, an eighteenth convolutional layer, a second Yolo prediction layer, a nineteenth convolutional layer, a second upsampling layer, an eighth route layer, a twentieth convolutional layer and a third Yolo prediction layer; wherein, The input layer, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the first route layer, the fifth convolutional layer, the second route layer, the sixth convolutional layer, the third route layer, the seventh convolutional layer, the fourth route layer, the eighth convolutional layer, the fifth route layer, the ninth convolutional layer, the sixth route layer, the third pooling layer, the tenth convolutional layer, the SPP layer, the eleventh convolutional layer, the fourth pooling layer, the twelfth convolutional layer, the fifth pooling layer, the thirteenth convolutional layer, the fourteenth convolutional layer, the fifteenth convolutional layer, and the first Yolo prediction layer are connected in sequence. The third convolutional layer is also connected to the first route layer, the second route layer, the third route layer, the fourth route layer, the fifth route layer, and the sixth route layer respectively. The thirteenth convolutional layer is also connected to the second Yolo prediction layer through the sixteenth convolutional layer, the first upsampling layer, the seventh route layer, the seventeenth convolutional layer, and the eighteenth convolutional layer. The eleventh convolutional layer is connected to the seventh routing layer and the nineteenth convolutional layer respectively, and the nineteenth convolutional layer is connected to the third Yoloprediction layer through the second upsampling layer, the eighth routing layer, and the twentieth convolutional layer; The input resolution of the input layer is 544×544.

7. The method for detecting ground objects by a drone according to claim 1, wherein: The Yolov3-tiny improved network Yolov3-tiny-PDB3 includes: an input layer, a first convolutional layer, a first pooling layer, a dense block, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a first route layer, a sixth convolutional layer, a second route layer, a first SPP layer, a seventh convolutional layer, a third pooling layer, a second SPP layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, a first Yolo prediction layer, an eleventh convolutional layer, a first upsampling layer, a third route layer, a twelfth convolutional layer, a thirteenth convolutional layer, a second Yolo prediction layer, a fourteenth convolutional layer, a second upsampling layer, a fourth route layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a seventeenth convolutional layer and a third Yolo prediction layer; wherein, The input layer, the first convolutional layer, the first pooling layer, the Denseblock, the second convolutional layer, the second pooling layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the first route layer, the sixth convolutional layer, the second route layer, the first SPP layer, the seventh convolutional layer, the third pooling layer, the second SPP layer, the eighth convolutional layer, the ninth convolutional layer, the tenth convolutional layer, and the first Yoloprediction layer are connected in sequence, the third convolutional layer is also connected to the second route layer, the fourth convolutional layer is also connected to the first route layer, the second route layer is also connected to the seventeenth convolutional layer, the seventh convolutional layer is also connected to the third route layer and the fourteenth convolutional layer respectively, the eighth convolutional layer is also connected to the second Yoloprediction layer in sequence through the eleventh convolutional layer, the first upsampling layer, the third route layer, the twelfth convolutional layer, the thirteenth convolutional layer and the second Yoloprediction layer. prediction layer; the fourteenth convolutional layer is connected to the third Yoloprediction layer through the second upsampling layer, the fourth routing layer, the fifteenth convolutional layer, and the sixteenth convolutional layer in sequence, and the seventeenth convolutional layer is also connected to the fourth routing layer; The input resolution of the input layer is 544×544.

8. The method for detecting ground objects by a drone according to claim 1, wherein: Under strong ambient light conditions, the Yolov3-tiny improved network Yolov3-tiny-VDB1 is constructed for ground vehicle detection; Under weak ambient light conditions, the Yolov3-tiny improved network Yolov3-tiny-VDB2 is constructed for ground vehicle detection; Under extremely weak ambient light conditions, the Yolov3-tiny improved network Yolov3-tiny-VDB3 is constructed for ground vehicle detection; Under strong ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB1 was constructed to detect ground personnel. Under weak ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB2 is constructed to detect ground personnel; Under extremely weak ambient light conditions, the improved Yolov3-tiny network Yolov3-tiny-PDB3 is constructed for ground personnel detection.

Images