Unmanned aerial vehicle detection method based on YOLOv8

Abstract

The invention discloses an unmanned aerial vehicle detection method based on YOLOv8, and the method comprises the steps: designing a multi-scale dense connection and a multi-layer shared detection head, and enabling the multi-scale dense connection to add a plurality of intermediate nodes in a weighted bidirectional feature pyramid network, and adding the connection between the nodes; and the multi-layer shared detection head performs shared convolution on the detection head of the YOLOv8 to balance the contribution of each detection head. According to the detection network YOLO-Drone of the small-target unmanned aerial vehicle in the ground-to-air scene, multi-scale dense connection and multi-layer shared detection heads are adopted in the network, the detection precision is effectively improved, redundancy of the detection heads is reduced, and the actual application requirement is met.

Description

Technical Field

[0001] The present invention belongs to the technical field of small target detection and relates to a drone detection method based on YOLOv8. Background Art

[0002] With the development of deep learning technology, vision-based detection methods have become a research focus in the field of drone detection due to their high accuracy, wide applicability, and reduced human intervention. For example, some researchers have used advanced object detection networks such as the YOLO series, the R-CNN series, SSD, and DETR for drone detection. Although deep learning methods have been widely used in drone detection, most research focuses on image or video processing, with relatively little attention paid to real-time stream processing and edge deployment. By summarizing the current research status in the introduction, the main challenges of real-time drone detection are summarized as follows:

[0003] (1) How to design an accurate and fast detection algorithm to cope with the small and fast target characteristics of drones.

[0004] (2) How to effectively deploy the algorithm on edge platforms whose computing power is far lower than that of servers. Summary of the Invention

[0005] To solve the above problems, the present invention provides a drone detection method based on YOLOv8, including designing multi-scale dense connections and multi-layer shared detection heads. The multi-scale dense connections add several intermediate nodes in a weighted bidirectional feature pyramid network and increase connections between the nodes. The multi-layer shared detection head performs shared convolution on the YOLOv8 detection head to balance the contribution of each detection head.

[0006] Preferably, the designing of multi-scale dense connections includes:

[0007] First, a round of upsampling and downsampling is performed after removing the backbone network in the weighted bidirectional feature pyramid network;

[0008] Secondly, we draw on the UNet++ structure, add several intermediate nodes to the network, and incorporate feature fusion steps;

[0009] Finally, dense connections are added between nodes.

[0010] Preferably, the design of a multi-layer shared detection head includes:

[0011] First, the convolutional path of regression is merged into a shared convolution operation in the detection head of YOLOv8;

[0012] Secondly, feature maps of different scales share weight convolution;

[0013] Finally, the number of channels of feature maps at each scale is unified to 256 to enhance the importance of shallow features in small object detection, balance the contributions of different feature layers, and reduce the number of parameters.

[0014] The beneficial effects of the present invention include at least: the YOLO-Drone algorithm proposed in the present invention significantly outperforms six other algorithms (including Faster R-CNN, RTMDet, YOLOv10n, etc.) in small target detection, demonstrating excellent accuracy. Its multi-scale dense connection improves P by 1%, R by 19.7%, mAP50 by 9.7%, and mAP50:95 by 6.1% compared to the baseline model; on this basis, the multi-layer shared detection head further improves mAP50:95 by 0.6%, while reducing the number of detection head parameters by more than 80%. In addition, the detection speed of YOLO-Drone is comparable to that of the lightweight YOLO series, only 0.1ms slower, which basically meets the requirements of high-precision real-time drone detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a diagram of the multi-scale dense connection process of the drone detection method based on YOLOv8 in the present invention;

[0016] Figure 2 This is a diagram of different connection modes used in ablation experiments for the drone detection method based on YOLOv8 of the present invention;

[0017] Figure 3 This is a structural diagram of the detection head of the YOLOv8n model of the YOLOv8-based drone detection method according to a specific embodiment of the present invention;

[0018] Figure 4 This is a multi-layer shared detection head diagram of the drone detection method based on YOLOv8 according to a specific embodiment of the present invention;

[0019] Figure 5 Two detection head structures of the YOLOv8n model in the prior art;

[0020] Figure 6 This is a YOLO-Drone network structure diagram of the drone detection method based on YOLOv8 in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0022] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0023] The present invention includes designing multi-scale dense connections and multi-layer shared detection heads, wherein the multi-scale dense connections add several intermediate nodes in the weighted bidirectional feature pyramid network and increase connections between the nodes; the multi-layer shared detection head performs shared convolution on the YOLOv8 detection head to balance the contribution of each detection head.

[0024] See also Figure 1 , the design of multi-scale dense connections is derived from the weighted bidirectional feature pyramid network (BiFPN,WeightedBi-directionalFeaturePyramidNetwork)BIFPN( Figure 1 (a) performs well in small object detection, but its computational efficiency is low on edge platforms due to its deeper network structure and weighted feature fusion method. Therefore, several reasonable modifications are made.

[0025] First, BIFPN performs three rounds of upsampling and downsampling after removing the backbone network, while YOLOv8 only performs one round. To this end, the redundant sampling stages are removed, such as Figure 1 (b) shown.

[0026] Secondly, by referring to the UNet++ structure, multiple intermediate nodes are added to the network ( Figure 1 (c)), and incorporate the necessary feature fusion steps ( Figure 1 (d));

[0027] Finally, to achieve more comprehensive feature fusion, dense connections are added between nodes, ultimately forming a multi-scale dense connection structure. This design not only improves the ability of feature extraction, but also effectively enhances the detection accuracy of small objects.

[0028] The core of designing multi-scale dense connections lies in its additional intermediate nodes, connection methods and dense connections. Figure 2 The best structure was obtained through subsequent experiments. Figure 2 In (a), (b) and (c), the nodes follow the UNet++ feature fusion and merge features along the sampling direction, which is called Fuse1; Figure 2 In (d), (e) and (f), the node merges features with the upper and lower layers to obtain richer information, which is called Fuse2.

[0029] As shown in Table 1, the ablation experiment shows that both Fuse1 and Fuse2 improve the performance of YOLOv8n, among which Fuse2 performs slightly better due to its richer feature fusion. Although Dense1 and Dense2 have no obvious advantages, Dense1 performs better when the fusion is not comprehensive, while Dense2 performs better when the feature fusion is richer. Figure 2 The multi-scale dense connections proposed in (f) achieve the best results, showing its effectiveness in small object detection.

[0030] Table 1 Multi-scale dense connection ablation experiment

[0031]

[0032] Regarding the design of a multi-layer shared detection head, by checking the printout of the YOLOv8n model, we can find that the detection head contains 751,507 parameters, while the entire model has 3,011,043 parameters, which means that the detection head accounts for about a quarter of the model. This difference is due to the different structures between YOLOv8 and YOLOv5. YOLOv8 uses a decoupled head structure, see Figure 3 , providing two independent convolutional paths for each scale of the output feature map: one for bounding box regression and the other for classification regression. However, in single-object detection tasks, the algorithm only needs to locate the target without worrying about classification. This decoupled convolution strategy leads to redundancy in the detection head. To improve efficiency, YOLO-Drone introduces a multi-layer shared head to reduce parameter redundancy and improve model performance.

[0033] Therefore, the present invention redesigns the detection head of YOLOv8 and proposes a multi-layer shared detection head. Figure 4 . First, the multi-layer shared detection head merges the convolution path of the regression into a shared convolution operation to reduce redundancy; second, feature maps of different scales share weight convolution, further optimizing the efficiency of the convolution operation. Finally, the number of channels of feature maps of each scale is unified to 256 to enhance the importance of shallow features in small object detection, thereby balancing the contributions of different feature layers and reducing the number of parameters. By comparison, the detection head of YOLOv8n+multi-scale dense connection contains 617,364 parameters, while the detection head of YOLOv8n+multi-scale dense connection+multi-layer shared detection head has only 109,461 parameters, a reduction of more than 80%. This result shows that the multi-layer shared detection head performs well in optimizing the number of parameters.

[0034] In order to effectively optimize the detection head of YOLOv8, the multi-layer shared detection head is compared with two other common detection head improvement methods, see Figure 5 . Figure 5 (a) Head1 is obtained by merging two convolution branches. Figure 5 (b) The previous 3×3 convolution is replaced by 1×1 convolution to obtain Head 2. Both detection heads can significantly reduce the amount of computation and the number of model parameters.

[0035] The results are shown in Table 2 (MDC: Multiscale Dense Connection, MSH: Multilayer Shared Head). Although both Head1 and Head2 reduce the computational burden of the detection head, the performance indicators drop significantly. In contrast, although the proposed multi-layer shared detection head shows a certain performance drop, it improves on key indicators such as mAP50:95 and maintains mAP50. This shows that for the characteristics of small object detection, the multi-layer shared detection head improves the expressiveness of the shallow detection head by sharing convolution parameters, thereby improving the overall performance. Although other layers of the model can be optimized through pruning, the detection head directly affects the detection results. Cautious pruning will greatly reduce the performance of the model. Therefore, it is crucial to improve the detection head structure to achieve lightweight.

[0036] Table 2 Performance of different detection heads

[0037]

[0038] See also Figure 6 This paper ultimately proposes a new detection network based on YOLOv8, named YOLO-Drone. This network integrates an innovative multi-scale dense connection module and a multi-layer shared detection head. The multi-scale dense connection module enhances feature extraction capabilities by adding intermediate nodes and skip connections, while the multi-layer shared detection head balances the contributions of each detection head through shared convolution, thereby improving the detection accuracy of small objects.

[0039] 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 and improvements 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 drone detection method based on YOLOv8, characterized in that: The method includes designing multi-scale dense connections and multi-layer shared detection heads. The multi-scale dense connections add several intermediate nodes in the weighted bidirectional feature pyramid network and increase connections between the nodes. The multi-layer shared detection head performs shared convolution on the YOLOv8 detection head to balance the contribution of each detection head.

2. A drone detection method based on YOLOv8 according to claim 1, characterized in that: The design of multi-scale dense connections includes: First, a round of upsampling and downsampling is performed after removing the backbone network in the weighted bidirectional feature pyramid network; Secondly, we draw on the UNet++ structure, add several intermediate nodes to the network, and incorporate feature fusion steps; Finally, dense connections are added between nodes.

3. The drone detection method based on YOLOv8 according to claim 1, characterized in that: The design of the multi-layer shared detection head includes: First, the convolutional path of regression is merged into a shared convolution operation in the detection head of YOLOv8; Secondly, feature maps of different scales share weight convolution; Finally, the number of channels of feature maps at each scale is unified to 256 to enhance the importance of shallow features in small object detection, balance the contributions of different feature layers, and reduce the number of parameters.

Images