Robust Tracking Occluded Human in Group by Perception Sensors Network System

Abstract

Tracking people even being partially or fully occluded in the group situation is studied using a Perception Sensor Network (PSN) system which is composed of multiple Kinects used to detect human 3D locations, and pan tilt zoom (PTZ) cameras used to identify human faces. A method is proposed to fuse multiple detection of human in the PSN system. After associating detected human with corresponding names, the novel grouping and ungrouping algorithms are proposed. When a group of multiple human staying close together is formed, viewpoint and illumination invariant features of group members including human 3D location, height, color and binary robust invariant scalable keypoint (BRISK), retrieved from region of interest (ROI) of both depth and color images, are then stored and updated into the group database. Based on the distance between a group location at previous frame and each member location in the group at current frame, the PSN system decides whether to keep the members in the group or to ungroup them then reassign the right name among the group database by minimizing multiple criterions. The experimental results demonstrate the outperforming of the proposed method on tracking people in group than conventional methods.