Photometric Invariant Visual Walking Stick Detection in Human-Robot Coexisting Environments: A Stratified Random Sampling Based Approach

This work focuses on solving a crucial problem within the genre of human-robot coexistence, that is, accurate detection of a walking stick held by the human target in the series of visual images acquired by the onboard vision sensor of a mobile robot, during people following. Considering that the stick poses are subjected to various transformations during the tracking process, this study demonstrates how the stick detection problem can be solved as a general-purpose template matching problem. A high-speed randomized template matching algorithm, namely FAsT-Match, which is designed for approximate template matching under 2D affine transformations, has been effectively implemented in our work of walking stick detection. However, the matching performance of the FAsT-Match degrades when the real-life stick images get affected by photometric variations. Moreover, keeping in mind the real-life applicability of the algorithm, only a minor fraction of pixels is scanned from the images at random for the evaluation of the fitness of individual transformations, and the degree of randomness influences the matching outcomes. The present study proposes an improved adaptation of the FAsT-Match for walking stick detection in photometrically affected human-robot coexisting environments, called stratified sampling based photometric invariant affine template matching (SS-PIATM). The proposed approach offers a robust solution in the face of photometric variations, as well as improves the matching accuracy by introducing a new stratified random-sampling based mechanism for the evaluation of affine transformations. Real-life performance evaluations aptly demonstrate the supremacy of the proposed SS-PIATM algorithm in walking stick detection during human following with a mobile robot. *