Online multi-person tracking using feature-less location measurements

Abstract

This thesis presents a scalable real-time multi-object tracking system based on feature-less location measurements. The thesis introduces a two-stage object tracking algorithm along with a server infrastructure that allows users to view the tracking results live, replay old frames, or compute long-term analytics based on the tracking results. In the first tracking stage, consecutive measurements are connected to form short tracklets using an algorithm based on MHT. In the second stage, the tracklets are connected to form longer tracks in an algorithm that reduces the tracking problem to a minimum-cost flow problem. The system infrastructure allows for a large number of connected devices or sensors while reducing the possible points of failure. The tracking algorithms are evaluated in a controlled environment and in a daylong experiment in a real setting. In the latter, the number of people detected by the tracking algorithms was correct 83% of the time when tracking was done using noisy motion-based measurements.