Year: 2025
Type: conference-paper
Abstract: In recent years, the increasing number of vehicles has led to heightened traffic congestion. Traditional surveillance is limited to monitoring fixed locations. However, unmanned aerial vehicle (UAV) photography provides advantages such as convenience and a wide perspective, gradually replacing these conventional techniques for traffic video collection. Nonetheless, challenges including complex scenes, varying scales and frequent occlusions result in issues such as false positives, missed targets, and identity switches (IDSW). To tackle these, we developed the UVD detection algorithm based on YOLOv8-n and the AByteTrack tracking algorithm derived from ByteTrack. The integration of these two components forms the UVD-ATrack algorithm. In the backbone of the detection section, we introduced the RFAConv and designed RF_C2f to enhance feature extraction capabilities. For the neck component, the adaptive weight fusion module (AWFM) was developed to effectively combine multi-scale local information. During the tracking phase, we optimized the motion model. Additionally, camera motion compensation (CMC) was implemented to adaptively mitigate the effects of sharp perspective changes. Test results utilizing the Visdrone dataset indicate that [email protected] and [email protected]:0.95 improved by 4.4% and 3.6% respectively in the detection section. In the tracking test, IDF1, HOTA, MOTA and MOTP also exhibited significant enhancement while IDSW decreased by 77%. Compared with state-of-the-art methods, our proposed algorithm demonstrates exceptional performance.
Authors: Chengxuan Li, Wei Yang, Chang Sun
Institution: Xihua University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411015893