Research Article
YOLOv5-LW: Lightweight UAV Object Detection Algorithm Based on YOLOv5
@INPROCEEDINGS{10.1007/978-3-031-55471-1_2, author={He Xiao and Kai Zhao and Xiaomei Xie and Peilong Song and Siwen Dong and Jiahui Yang}, title={YOLOv5-LW: Lightweight UAV Object Detection Algorithm Based on YOLOv5}, proceedings={Mobile Networks and Management. 13th EAI International Conference, MONAMI 2023, Yingtan, China, October 27-29, 2023, Proceedings}, proceedings_a={MONAMI}, year={2024}, month={3}, keywords={Object detection UAV object detection Lightweight model Small object detection}, doi={10.1007/978-3-031-55471-1_2} }- He Xiao
Kai Zhao
Xiaomei Xie
Peilong Song
Siwen Dong
Jiahui Yang
Year: 2024
YOLOv5-LW: Lightweight UAV Object Detection Algorithm Based on YOLOv5
MONAMI
Springer
DOI: 10.1007/978-3-031-55471-1_2
Abstract
UAV object detection task is a highly popular computer vision task, where algorithms can be deployed on unmanned aerial vehicles (UAVs) for real-time object detection. However, YOLOv5’s performance for UAV object detection is not entirely satisfactory due to the small size of the detected objects and the problem of occlusion. To address these two issues in the YOLOv5 algorithm, we propose the YOLOv5-LW algorithm model. Building upon YOLOv5, we replace the FPN-PAN network structure with the FPN-PANS structure. This modification helps mitigate the issue of feature disappearance for small objects during the training process while reducing the model parameters and computational complexity. Additionally, within the FPN-PANS structure, we employ a multistage feature fusion approach instead of the original feature fusion module. This approach effectively corrects the erroneous information generated during the upsampling stage for certain objects. Finally, we replace the SPPF module with the SPPF-W module to further increase the receptive field while maintaining almost unchanged parameters. We conducted multiple experiments and demonstrate that YOLOv5-LW performs exceptionally well in lightweight small object detection tasks using the VisDrone dataset. Compared to YOLOv5, YOLOv5-LW achieves a 4.7% improvement in mean average precision (mAP), reduces the model size by 40%, and decreases the parameters by 40%.
