Experimental Tests of Vision-Based Artificial Landmark Detection using Random Forests and Particle Filter

Abstract

This paper proposes a novel artificial landmark detection technique for underwater robots in structured underwater environment. The novel landmark detection technique is composed of a salient object segmentation using random forest combined with particle filter and an object recognition using weighted template matching. The random image patch-based random forest is employed for detection of the regions of salient ocjects and its accuracy is enhanced by cimbining with particle filter. Each detected candidate region is refined through the active contour technique and recognized as one of the artificial landmark or background by the weighted template matching technique. The performance of the proposed method is evaluated by experiments with an autonomous underwater robot platform, yShark, developed by KRISO and the results are discussed by comparing with the results of the previous research. with particle filter and an object recognition using weighted template matching. The random image patch-based random forest is employed for detection of the regions of salient ocjects and its accuracy is enhanced by cimbining with particle filter. Each detected candidate region is refined through the active contour technique and recognized as one of the artificial landmark or background by the weighted template matching technique. The performance of the proposed method is evaluated by experiments with an autonomous underwater robot platform, yShark, developed by KRISO and the results are discussed by comparing with the results of the previous research.