• Autonomous
• Detection
• Navigation
• Obstacle

The work focuses on an obstacle detection framework for an autonomous robotics platform. State of the art classical and Deep Learning existing frameworks and approaches are first explored. A probabilistic framework that includes 3D reconstruction techniques, stereo matching and UV-disparity projection is designed. The system is integrated with appearance cues, edges related metrics and ground plane homography estimation to robustly segment obstacles in a wide range of scenarios. A synthetic and real datasets are generated to test the performances and accuracy of the proposed algorithms. The framework is deployed on a autonomous mobile robot where real time performances are achieved, using CUDA computing capabilities of a NVIDIA Xavier embedded platform.