STRACK: Robust Tracking of Small Objects in Low-Light Conditions

Tracking small objects in low-light conditions presents significant challenges for multi-object tracking (MOT) methods due to diminished image quality, motion blur, poor contrast, and increased noise levels. This research presents STRACK, a robust MOT framework specifically designed to address these challenges through the effective integration of advanced detection, motion modeling, and tracklet association methods. STRACK uses a fine-tuned YOLOv11 detector, ensuring reliable detection across varying object sizes and illumination conditions. The framework proposes an adaptive noise-aware Enhanced Kalman Filter (EKF) for reliable motion prediction, transformer-based Dynamic Motion Compensation (DMC) for reducing camera motion effects, and a Spatio-Temporal Relational Network (STRN), which combines LSTM and graph convolutional modules to strengthen associations under appearance degradation. Additionally, STRACK uses Adaptive Motion Smoothing (AMS) with Gaussian process regression to address missed detections, alongside IoU-ReID fusion to combine motion and appearance signals. Extensive experiments on the MOT17 and MOT20 benchmarks indicate that STRACK significantly surpasses leading methods, achieving HOTA scores of 66.71 and 66.44, respectively, while maintaining real-time performance at 25 FPS. These results confirm STRACK as a reliable and scalable platform for robust MOT in challenging low-light environments for applications in autonomous driving, surveillance, and ecological monitoring.