Towards unsupervised canine posture classification via depth shadow detection and infrared reconstruction for improved image segmentation accuracy

Abstract

Hardware capable of 3D sensing, such as the Microsoft Kinect, has opened up new possibilities for low-cost computer vision applications. In this paper, we take the first steps towards unsupervised canine posture classification by presenting an algorithm to perform canine-background segmentation, using depth shadows and infrared data for increased accuracy. We report on two experiments to show that the algorithm can operate at various distances and heights, and examine how that effects its accuracy. We also perform a third experiment to show that the output of the algorithm can be used for k-means clustering, resulting in accurate clusters 83% of the time without any preprocessing and when the segmentation algorithm is at least 90% accurate.