Choice Of Distance Metrics in DBSCAN Based Color Template Matching Applied to Real-Time Human Shoe Detection

Abstract

In human-robot collaborative environments, human subject detection and tracking is one of the most pertinent problems in recent times. In some of our recent works, we have demonstrated how this problem can be addressed from a vision sensor-based perspective, by utilizing general-purpose template matching algorithms for the purpose. A state-of-the-art such algorithm, namely the FAsT-Match, and its improved variant for RGB color images, termed the CFAsT-Match, can be successfully implemented in real robots for the purposes of visual human shoe detection, during people following. The CFAsT-Match involves the use of a popular density-based clustering algorithm, named DBSCAN, to form irregular-shaped clusters of the template image pixels. In this paper, we have presented a detailed study, where we implement various distance metrics while clustering the template image using the DBSCAN algorithm, and investigate the effects on the final detection outcomes.