Combining Distance Measures to Classify Terrain Image Sequence Based on Dynamic Texture Model

Abstract

Utilizing image sequences obtained from onboard cameras is important to improve autonomous mobility for planetary rovers. In this paper, we propose a method based on a linear dynamical system model called a Dynamic Texture to classify terrain image sequences which contain different types of properties. Among some physical properties included in image sequences, we focus on static properties such as a soil texture and dynamic properties such as a constant image velocity. Distance measures based on an observability space and a time-series model called an autoregressive (AR) model are used for classifying the properties. A cross-validation test using real image sequences shows that the latter distance measures are more advantageous than the former measures with respect to the dynamic properties. Thus we combine two distance measures specific to different properties to gain the classification performance. Experimental results demonstrate the effectiveness of combining distance measures.