Trajectory-guided detection and feature reuse for infrared object tracking based on YOLOv8–DeepSORT

Abstract

Infrared object tracking is widely used in intelligent surveillance and autonomous perception systems due to its robustness under low-illumination and adverse environmental conditions. However, conventional YOLOv8–DeepSORT frameworks exhibit limited performance in infrared scenarios because of weak thermal contrast, background clutter, unreliable appearance representations, and long-term thermal occlusion. To address these issues, this paper proposes a lightweight trajectory-guided detection and feature reuse framework for infrared object tracking based on YOLOv8– DeepSORT. First, a trajectory-guided dual-threshold detection strategy is introduced to improve weak target recall while suppressing thermal clutter interference. Second, deep semantic features extracted from YOLOv8 are directly reused for target association, replacing visible-light pretrained appearance embeddings in DeepSORT. In addition, an adaptive trajectory retention strategy is designed to improve tracking continuity during long-term thermal occlusion. Experimental results on the LSOTB-TIR dataset demonstrate that the proposed method significantly improves identity consistency and tracking robustness compared with the baseline YOLOv8–DeepSORT framework. Specifically, the proposed method improves IDF1 from 16.48% to 59.93% and MOTA from -4.15% to 21.62%, while reducing identity switches to zero. The proposed framework does not require additional infrared ReID training or network modification, making it suitable for practical real-time infrared tracking applications.