Multi-pedestrian tracking for far-infrared pedestrian detection on-board using particle filter

Target tracking is important for pedestrian detection in on-board vision application preventing traffic accidents effectively. Facing complex traffic scene including background change, various pedestrian appearance and multi-targets etc., existing target tracking algorithms such as Kalman and particle filters expose shortcomings in accuracy, robustness and availability. This paper proposes an improved particle filter algorithm for multi-target tracking in far-infrared (FIR) pedestrian detection, where a heuristic tracking scheme including feature model learning and target tracking iteratively is used. Partial least squares regression (PLSR) and heuristic computation are adopted to learn and update feature models for each pedestrian. The proposed particle filter algorithm combines adaptive searching region and double feature models, to achieve higher target tracking performance. Experiment on several FIR video sequences demonstrates the improved scheme outperforms comparing with other particle filter algorithms when multi-pedestrian tracking, even with partial occlusion, scale and posture variation.