| Abstract |
Tracking multiple people simultaneously in real-world environments is crucial for a wide variety of applications, including video surveillance and human-robot interaction. In this work, we present a method for tracking an unknown number of humans from a robotic platform. This is accomplished by using the robot's 2D Laser-Range Scanner sensor located at the feet height and a plain web camera. Since we employ two separate sensory modalities, we commence by providing a method to manually, yet with minimal effort, calibrate the sensory setup.
The proposed tracking methodology is an iterative process which consists of a clustering and a tracking step. During the first step, human centers are being produced from the fusion of sensory data by means of clustering. The developed clustering method is based on the newly introduced Human Evidence Grid (HEG). HEG constitutes a 2D Cartesian occupancy grid, whose cells capture the occupancy evidence -i.e. a detected human face on the image plane occupies the respective cell- of environment targets, being monitored by the LRS scanner. The occupancy likelihood, that the HEG holds, is calculated according to a training process. Several training datasets were considered in order to robustly learn real-life configurations among a human face and the corresponding LRS measurements that the detected face may produce. The potential of the developed methodology is explored by comparing it to a Euclidean approach which considers only LRS data. The results clearly indicate that HEG clustering outperforms the Euclidean approach. In the second step, the produced cluster centers are subsequently being tracked by a Rao-Blackwellized Particle Filtering (RBPF) Multiple Target Tracker, where human appearance and disappearance are coded via a distribution that models the lifetime of each target.
Several datasets were captured in order to test the proposed approach in real-life scenarios. The results from a dataset, where three people are being monitored and tracked in the surveillance area, are presented and described in detail. Our experimental results demonstrate the effectiveness of HEG clustering in handling temporary occlusions in the LRS area.
|
Simultaneous tracking of multiple moving humans from a robotic system
