Visual Detection for Robotic Person Following in Environments with Significant Light Variations

Abstract

The paper proposes a vision based robotics person following system that is robust with respect to significant variations of lighting and person’s poses and outfits. It can follow a target person when he makes challenging motions such as jumping from side to side and in the presence of other persons. In addition to the integration of methods such as stereo block matching, histogram of oriented gradients and head detection, several new the techniques are developed in this paper. These include methods to identify persons when they disappear from the camera view and then reappear, and reacquiring the target when she/he is lost. Furthermore, an adaptation of the trained feature vector is proposed to make the process insensitive to environmental light changes. An important aspect of the system is the use of inexpensive and readily available components which makes it a viable solution. Numerous trials have been conducted to demonstrate the performance of the system to significant changes