Hierarchical Supervised Network for Multi-Object Tracking in Reconnaissance UAV Video under Foggy Weather

Multi-object tracking in reconnaissance UAV videos under foggy weather presents several challenges: (1) the boundaries between the target and the background become blurred; (2) targets have limited target pixels and few available features; (3) reconnaissance UAV movement causes variations on shape and size of targets. We tackled these challenges by developing a Hierarchical Supervised Network for multi-object Tracking (HSTrack). First, we employed the Salient Foreground Perception (SFP) module to extract foreground targets from blurry backgrounds. SFP is incorporated to enhance the limited pixel objects and increase foreground-background contrast. Second, we utilized multi-scale object detection (MSOD) to predict the objects with scale variations. Meanwhile, to implement hierarchical supervision, we devised a unified loss function to compute the loss of the prediction between MSOD and SFP. Finally, we use mixed metrics to connect detections to tracklets, achieving multi-object tracking. We validated our approach through experiments on a self-assembled dataset collected from a reconnaissance UAV under foggy weather, named UAV-Fog. The experimental results demonstrated that our proposed method outperformed other state-of-the-art methods in reconnaissance UAV videos under foggy weather. Furthermore, we validated the effectiveness of HSTrack on the VisDrone2019-MOT and UAVDT datasets, where it also achieved outstanding performance.