Depth image based terrain recognition for supervisory control of a hybrid quadruped

Abstract

This paper presents the depth image based locomotion strategy selection framework for a hybrid mobile robot. Terrain recognizer is a major component of a supervisory controller which classifies depth images into terrain types in real-time and selects different locomotion mode sub-controllers. In order to design the terrain recognizer, a database consisting of five terrain types (uneven, level ground, stair up, stair down and not traversable) is generated. Confidence based filtering is applied to enhance depth image data. The accuracy of the terrain classification for the testing database in five class terrain recognition problem is 96.71%. Real-world experiments conducted in mixed terrain environment evaluate both locomotion and terrain recognition capabilities of the robot in real-time. Experimental results show that a consumer depth camera might serve as an effective instrument for terrain recognition and thus locomotion strategy selection for hybrid robots with multiple locomotion modes.