Real Time People Tracking in Crowded Environments with Range Measurements

Abstract

Social and assistive robots have recognised benefit for future patient care and elderly management. For real-life applications, these robots often navigate within crowded environments. One of the basic requirements is to detect how people move within the scene and what is the general pattern of their dynamics. Laser range sensors have been applied for people tracking in many applications, as they are more precise, robust to lighting conditions and have broader field of view compared to colour or depth cameras. However, in crowded environments they are prone to environmental noise and can produce a high false positive rate for people detection. The purpose of this paper is to propose a robust method for tracking people in crowded environments based on a laser range sensor. The main contribution of the paper is the development of an enhanced Probability Hypothesis Density (PHD) filter for accurate tracking of multiple people in crowded environments. Different object detection modules are proposed for track initialisation and people tracking. This separation reduces the misdetection rate while increasing the tracking accuracy. Targets are initialised using a people detector module, which provides a good estimation of where people are located. Each person is then tracked using different object detection module with a high accuracy. The state of each person is then updated by the PHD filter. The proposed approach was tested with challenging datasets, showing an increase in performance using two metrics.