Concatenating multiple trajectories using Kalman filter for pedestrian tracking

This work proposes a tracking-by-detection methodology for pedestrians following using a targlet framework. In this framework each pedestrian is considered as an autonomous agent, denominated targlets, and modeled with a state machine. The tracking procedure is initialized by a people detector computed on a Movement Feature Space. Detector outputs generate probabilistic fields employed for the tracking of each targlet, in order to obtain their individual trajectories along the sequence. The targlet framework is then analyzed off-line to: filter false positives, and concatenate broken trajectories of the same person using a Kalman filter approach. The system is tested on the public dataset PETS2009 S2.L1 obtaining good results, similar to the best methodologies in the state of the art.