Academic Report: From Autonomous Motorcycle to Bridge Deck Scanning: Visual Navigation for Size and Power Constrained Mobile Robots
Speaker: Dezhen Song
Time:2017-7-28 15:00-16:00
Location：The Meeting Room 304 of Hydraulic Building in Yuquan
Abstract:
In this talk, I report our 14 years’ research in algorithms and systems designed to assist size and power limited robots/vehicles to perform self-navigation in both outdoor and indoor environments, and their applications ranging from self-driving cars to bridge deck inspection. I will begin with how we develop the world’s first autonomous motorcycle and our entrance to Darpa Grand Challenges 2004 and 2005, where we design an appearance-based approach for the autonomous motorcycle to navigate in desert terrain. For GPS-challenged environments, we investigate different visual features to improve robustness and speed for simultaneous localization and mapping (SLAM) algorithms. We develop high-level landmark-based SLAM algorithms to help low cost robots with a monocular camera to obtain the same level localization accuracy as the robot with sophisticated sensors. To handling large lighting variations, we also develop a point and line-based approach visual odometry method for robots with RGB-D cameras. At the end, we will discuss our recent development on a featureless motion-vector based SLAM and its application in bridge deck scanning. We employ a sensor fusion approach that combines visual SLAM with ground penetration radar, lidar, and GPS inputs to map subsurface and surface cracks in the bridge deck for freeway maintenance.