A Fast Method for Tracking People with Multiple Cameras

Abstract

We propose a multi-camera method to track several persons using constraints from the epipolar and projective geometries. The method is very accurate, fast, and simple. We first compute accumulator images for each time frame that shows the probability of object positions on the ground. We developed a voting based method that allows employment of the integral images to make the accumulator computation very fast. Next, we perform two-pass 3D tracking on the volume generated by stacking these accumulator images. Our main contributions are the fast computation of the accumulator images and application of fast 3D tracking methods like the Kalman Smoother instead of the computationally expensive methods like the Viterbi algorithm. The proposed tracking method is evaluated on people videos captured using four synchronized cameras.