Implementation of Vision-Based Navigation for Pedestrian Environments

Abstract

Autonomous navigation has rapidly grown to become a predominant field of study utilizing recent advances in robotics and artificial intelligence. Most autonomous navigation methods rely on expensive and complex sensor arrays such as Lidar, which pose practical limitations on the widespread deployment of these devices. This thesis presents an end-to-end implementation of a vision-based navigation pipeline for autonomous navigation in pedestrian environments, utilizing only a single front-facing RGB-D camera and tracking camera as perception devices. This pipeline utilizes 3D monocular object tracking in combination with an advanced Kalman-filter based geometric tracking scheme to track nearby pedestrians, in combination with full SLAM for localization and a reinforcement-learning based navigation stack to navigate through challenging dynamic multi-agent environments. The functionality of this pipeline is demonstrated through a series of pedestrian tracking and navigation experiments with many pedestrians. The tracking module of this pipeline is able to correctly localize pedestrians within 0.4 meters in simple scenarios and 0.6 meters in challenging multi-pedestrian stress testing cases inside of a 12 meter space in spite of limited field of view and relying on only inexpensive RGB camera images. Full end-to-end navigation was demonstrated in a crowded environment with 5 pedestrians, with only one collision out of 13 trials.