An Experimental Comparison of ROS-compatible Stereo Visual SLAM Methods for Planetary Rovers

Ego-motion estimation and awareness of the surroundings are crucial tasks to perform for an exploration vehicle. Visual SLAM techniques allow to fulfill these needs by extracting meaningful information from images: changes in the appearance of the environments are used to simultaneously estimate both its structure and the relative motion of the camera. While tipically heavyweight and computationally complex, thanks to the most recent advances in embedded platform architecture, most of the available open source Visual SLAM software can be run online on ARM processors. This, together with miniaturization and lower power consumption, opens great scenarios for autonomous navigation of mobile robots. This paper covers the implementation of most of the available open source Visual Stereo SLAM on a nVidia Jetson TX2 platform highlighting critical information such as pose estimation accuracy, capability of loop closure, CPU and memory usage. The measurement setup is mounted on the MORPHEUS (Mars Operative Rover of Padova Engineering University Students) rover and LiDAR data is used as reference for performance evaluation.