UAVs have become extensively used in various fields due to the rapid advancement of UAV control technology. Accurate detection and identification of objects around UAVs have become crucial during UAV applications. However, the high navigation characteristics of UAVs pose challenges in model feature extraction and object detection, given the varying scale of objects and the presence of numerous small maneuvering objects. Additionally, high-speed and high-altitude flight introduces motion blur, further complicating object differentiation. To address these challenges, this paper proposes the CSPCM-YOLOv5 algorithm. It includes a detection head for specifically detecting small objects. Part of the original C3 module of YOLOv5 is replaced with Cross Stage Partial ConvMixer (CSPCM) to enhance feature extraction. The original SPPF is replaced with SimCSPSPPF to improve feature expression. Furthermore, by integrating Coordinate Attention (CA) in front of the detection head, the network can effectively identify object regions in scenes with dense objects. The CIoU loss function is replaced with NWD Loss to enhance the detection performance for small objects. The improved CSPCM-YOLOv5 algorithm is evaluated on the VisDrone2019-DET validation set, demonstrating its effectiveness with significant performance improvements.
Research on UAV Object Detection Algorithm Based on CSPCM-YOLOv5
Lect. Notes Electrical Eng.
International Conference on Autonomous Unmanned Systems ; 2023 ; Nanjing, China September 09, 2023 - September 11, 2023
Proceedings of 3rd 2023 International Conference on Autonomous Unmanned Systems (3rd ICAUS 2023) ; Kapitel : 12 ; 119-128
27.04.2024
10 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch