Temporal-Aware Lightweight Visual Tracking Method for Dynamic Traffic Scenes

In traffic-based tracking scenes, acknowledging category diversity and visual field motion is pivotal. This leads tracking models to prioritize multi-category feature learning, motion analysis, and extended time series modeling rather than singularly focusing on attribute acquisition and background modeling. With this goal in mind, we construct a tracking model by integrating re-identification branches into a lightweight detection model. Simultaneously, we establish an inter-frame spatio-temporal attention mechanism using the Feature Transfer Module. This mechanism aligns the current frame features with the corresponding local features of the previous frame, which can bolster feature representation and downstream task performance. Experimental results on the traffic dataset validate the efficacy of the proposed approach in addressing challenges such as occlusion and deformation within traffic scenes. The proposed method attains a significant 31.5 MOTA and 35.7 AP on the BDD100K dataset with random data selection.