System for multiple people identification and tracking in video

Abstract

In this paper, a software system is presented that supports finding, tagging, identifying and tracking multiple people in videos with uncontrolled capturing conditions. The work was focused on two aspects. One is to build a parallel video processing pipeline that integrates image analysis modules such as face detection, recognition and tracking, efficiently and smoothly, so that multiple people can be simultaneously tracked in real time. Another aspect is to make innovations to each of the major image processing modules so that they are both fast and robust to variations in pose, illumination, occlusions and so on. Written in C++, this demo runs on a mainstream laptop. It can instantly recognize and constantly track multiple subjects in live or recorded videos.