Pedestrian Detection and Identification using Two Cameras

This paper describes a method for pedestrian detection, identification and tracking using image information. The method makes use of two cameras with a shared field of view and is robust to changes in illumination and shadows. After a brief calibration process, in which the scene is divided coarsely into planar pieces (which are later optimised), the process requires no interaction and automatically compensates for pairs of cameras with very different optical properties. Individual pedestrians are identified by the novel application of a process similar to the Wiener filter, which we call the regularised inverse filter. Experiments on outdoor scenes demonstrate that the method is robust to changes of illumination and shadows, successfully tracking over 9 out of every 10 pedestrians in challenging conditions.