How to tune a random forest for real-time segmentation in safe human-robot collaboration?

Abstract

This paper is an extension of our work related to a generic classification approach for low-level human body-parts segmentation in RGB-D data. In this paper, we discuss the impact of decision tree parameters, number of training frames and pixel count per object-class during a random forests classifier training. From the evaluation, we observed that a varied non-redundant training samples makes the decision tree learn the most. Pixel count per object-class should be just adequate otherwise it may lead to under/over-fitting problem. We found a highly optimized and a most optimal parameter setup for a random forests classifier training. Our new dataset of RGB-D data of human body-parts and industrial-grade components is publicly available for lease for academic and research purposes.