| "Autonomous vehicles (AV)" are not a new concept. They have been proposed for the long-term goal of having vehicles driving autonomously in an unknown environment for the purpose of human safety, convenience, and ease of life; both in civilian and military applications. In day-to-day life, humans take such navigational capabilities for granted. The human brain achieves all this simultaneously using mostly the sense of vision. That being one of the strong motivation behind vision based control, it is also desirable to use computer vision systems to achieve low cost, low maintenance autonomous vehicle navigation. Recent results from vision research suggest that visual control may now be feasible and will be a major step in conversion of the sensor system required for autonomous guidance to a vision-based one. In this dissertation, we begin the development of the component capabilities in Computer Vision, for autonomous road navigation for the long-term goal of having autonomous vehicles driving in an unknown environment. This dissertation limits itself to the monocular vision component of sensing and navigation control, and treats the path planning components superficially. Unlike other approaches in literature, this research work presents to achieve a complete vision based autonomous vehicle control reducing the drastic dependency on other sensors.;Proposed is a vision solution based on accumulator based parametric transforms to detect navigable regions in the urban environment. The novel accumulator voting scheme is called "Parametric Transform for Lanes (PTL)," which enables detection of multiple lanes and variations of lanes such as exits and intersections possible concurrently. Other than being robust to shadows, invariant to color, width and texture of the road, this PTL is based on a control-motivated philosophy. Thus, this dissertation proposes a novel total image-based control solution to the autonomous vehicle control problem, called "Predictive Control in Image Plane (PCIP)." PTL and PCIP together inherently support control for multiple operations; such as lane following, lane changing, turning at intersections, etc. Thus, the proposed predictive control not only reduces the control complexity but also increases the application space for vehicle control.;Research work is also extended to other visual components of autonomous driving such as sign detection, car detection, light detection, cross-walk detection etc. A novel clustering method is proposed for car backlight detection, as well as cross-walk detection. The dissertation explains a complete architecture for the autonomous urban driving and talks about integrating all these modules to achieve smooth control of the AV. |
