Obstacle Avoidance Based on Scene Depth Perception using Stereo Cameras for Autonomous Vehicles

Abstract

In this thesis, stereo vision has been explored in the context of autonomous vehicles, for obstacle avoidance. Stereo vision involves combining 2D views for 3D depth estimation, similar to how human vision works. The emphasis is on real-time algorithms for efficient obstacle tracking and navigation. Some of the depth perception algorithms discussed include stereo matching and semiglobal methods. Furthermore, for insights into pixel-level comparison of disparity maps, methodologies such as area based disparity, feature based disparity and triangulation have been covered. Mobile Robot Navigation strategies, categorized into seven types, are presented, including odometry, inertial navigation, magnetic compasses, active beacons, GPS, landmark navigation, and map-based positioning. Real-time obstacle avoidance algorithms using depth maps have been analyzed for ideal conditions and potential limitations. This thesis also includes a section on the design considerations of a four-wheel steering mobile robot platform. It explores adaptive steering control algorithms for optimal performance during manual operation, addressing challenges associated with four wheel steering mechanisms. This project thesis flow work is divided into 3 parts. The first being on ZED Stereo Camera for stereo vision for making of disparity maps followed by depth maps which then is used to estimate real time depth of the objects. The second is object detection and obstacle avoidance algorithm generation. The object detection models such as Yolo V4 lite is used to identify the obstacle in the path and then generated algorithm for obstacle avoidance is used as discussed further in the thesis. The final part is Mobile Base platform design and integration. A steering controlled mobile base is manufactured for real life testing if the algorithms in real time followed by the embedded integration using Jetson Nano