An improved YOLOv8s-based UAV target detection algorithm

Abstract

At present, the low-altitude economy is booming, and the application of drones has shown explosive growth, injecting new vitality into economic development. UAVs will face complex environmental perception and security risks when operating in low airspace. Accurate target detection technology has become a key support to ensure the orderly operation of UAVs. This paper studies UAV target detection algorithm based on deep learning, in order to improve detection accuracy and speed, and meet the needs of UAV autonomous perception under the background of low altitude economy. This study focuses on the limitations of the YOLOv8s target detection algorithm, including its low efficiency in multi-scale feature processing and insufficient small target detection capability, which hinder its ability to perform rapid and accurate large-scale searches for drones. An improved target detection algorithm is proposed to address these issues. The algorithm introduces AKConv into the C2F module. AKConv allows for convolution kernels with arbitrary numbers and sampling shapes, enabling convolution operations to more precisely adapt to targets at different locations, thereby achieving more efficient feature extraction. To further enhance the model's ability to extract critical features of small targets, the SPPF module incorporates the LSKA mechanism. This mechanism captures long-range dependencies and adaptivity more effectively while addressing computational complexity issues associated with large convolution kernels. Finally, the Bi-FPN feature pyramid network structure is introduced at the 18th layer of the model to accelerate and enrich feature fusion in the neck. Combined with the SCDown structure, a novel Bi-SCDown-FPN feature pyramid network structure is proposed, making it more suitable for detecting targets with insufficient feature capture in complex environments. Experimental results on the VisDrone2019 UAV dataset show that the improved algorithm achieves a 5.9%, 4.5%, and 6.1% increase in detection precision, detection recall, and mean average precision, respectively, compared to the original algorithm. Moreover, the parameter count and weight file size are reduced by 13.41% and 13.33%, respectively. Compared to other mainstream target detection algorithms, the proposed method demonstrates certain advantages. In summary, the target detection algorithm proposed in this paper achieves a dual improvement in model lightweighting and detection accuracy.

Fig 1. Structure diagram of the optimized YOLOv8s model.

The three improved parts are C2F_AKConv in orange, SPPF_LSKA in pink and SCDown in blue, and the blue line segment is the Bi-FPN feature pyramid network structure.

Fig 2. Convolution kernel of any size.

Fig 3. Convolution kernel of size 5 with arbitrary shape.

Fig 4. AKConv module structure.

Fig 5. LKA Network Structure.

Fig 6. LSKA Network Structure.

Fig 7. BiFPN Structure Diagram. There are 5 levels of nodes from P3 to P7.

The straight line represents the feature transmission path, and the orange arc represents the bidirectional connection relationship, showing the flow and fusion of features between different levels.

Fig 8. Bi-SCDown-FPN Structure Diagram.

Fig 9. SCDown Module Structure Diagram.

Fig 10. Target Sample Statistics of VisDrone2019.

Fig 11. Partial Training Images of the Dataset.

Fig 12. Improved training results of YOLOv8s.

Fig 13. mAP50 comparison graph before and after improvement.

Fig 14. Recall rate comparison graph before and after improvement.

Fig 15. Detection results of the original YOLOv8s in open space.

Fig 16. Detection results of the improved YOLOv8s in open space.

Fig 17. Detection results of the original YOLOv8s on densely parked roads.

Fig 18. Detection results of the improved YOLOv8s on densely parked roads.

Fig 19. Detection results of the original YOLOv8s at low light intersections.

Fig 20. Detection results of the improved YOLOv8s at low light intersections.

Fig 21. Comparison diagram of mAP50.

Fig 22. Comparison diagram of detection accuracy.

Fig 23. YOLOv3-spp.

Fig 24. YOLOv5s.

Fig 25. YOLOv7tiny.

Fig 26. Improved YOLOv8s.

Fig 27. YOLOv9c.

Fig 28. YOLOv10s.