Vision Sensing based People Following Robot: A Superpixel Augmented Density based Clustering Approach

This paper presents a new approach to vision based human shoe detection for its effective implementation in leader-follower modules in human-robot coexisting environments. During motion tracking, the poses of the target shoes may undergo various general transformations such as scales, rotations 2D translations, or 2D affine transformations. Thus, the shoe detection problem can be solved as a template matching problem under general conditions. Recently, a general-purpose contemporary template matching algorithm, called FAsT-Match, and an extended variant of it for RGB color domain, called CFAsT -Match, have been successfully implemented for shoe detection in successive frames captured during real-life human tracking. CFAsT-Match uses a density based clustering algorithm, namely DBSCAN, to create arbitrarily shaped clusters of the pixels of a specific template shoe image. This cluster information is required to map the color distribution from the template to the target shoe images (frames). In the present study, an advanced superpixel augmented DBSCAN based shoe detection approach is proposed, where a superpixel generation scheme in cascade with DBSCAN clustering is integrated into the CFAsT-Match for the purposes of shoe detection. Superpixel segmentation offers grouping of image pixels into perceptually meaningful entities that are more congruent with the visual system of a human being and that can act as primitives for further computation. The resulting configuration not only significantly reduces the computational complexity of the shoe detection framework, but also improves the detection accuracy. Real-life performance evaluation aptly demonstrates the supremacy of the proposed approach combining superpixel-segmentation and density based clustering.