Deep layer nested aggregation: anti-occlusion network for UAV detection and tracking
Jun Ma, Longchao Li, Shilin Huang, Xuzhe Wang, Yan Ping Guo, Dongyang Jin · 2024
Detecting and tracking the unmanned aerial vehicle (UAV) requires the ability of convolutional neural networks to extract identity features as much as possible. However, it is a challenge to detect small-sized UAVs because of limited available features by using the conventional networks. Additionally, complex occluded scenes, such as scale variations, lighting conditions, and weather changes, will affect the performance of detecting and tracking UAVs at low-altitude. Although some progresses have been made, the failure in detection and tracking for low-altitude UAVs often occurs due to the complex background occlusion. To ensure the consistency in the identification of UAVs flying through occluded scenes and enhance the ability of detection and tracking, we proposed a deep layer nested aggregation network architecture (DLNA) based on DLA-34 and the FairMOT framework for detecting and tracking the UAVs in occluded scenes and achieved a tracking accuracy increased by 16.1% compare to the conventional networks. Experimental results verify the feasibility and effectiveness of the DLNA for low-altitude UAV detecting and tracking in various occluded scenes.