RGB-D People Detection and Tracking from Small-Footprint Ground Robots

Abstract

Reliably detecting and tracking people from a low-lying viewpoint is a challenging problem, due to the uncommon viewpoint, from which only few features are visible. This paper presents a lower part-based approach, which enables to detect and track people in indoor environment from a small-footprint ground robot using a RGB-D sensor. The approach is based on the fact that the lower part of the human body is the most prominent feature from a low-lying viewpoint. By removing the ground plane firstly, the objects on it within a limited height are clustered. And their HOG features are computed and then fed to a pre-trained SVM binary soft classifier. The clusters with high HOG confidence are classified as people, and then they are performed as input of the tracking module. A two-term likelihood consisting of color coherence and distance coherence is exploited for the data association process for looking for the correct tracks. Tests have been performed in three different indoor environments. The experimental results show that the precision, recall and f1-score, which are the people detection metrics reaches 0.82, 0.78, and 0.80 respectively.