Autonomous vehicles often rely on highly accurate map data for navigation purposes. However, such maps are of limited availability and restrict the vehicle’s region of operation to already mapped areas. This thesis describes a vision-only perception system for autonomous obstacle avoidance and route following in unknown environments. It works without the need for global positioning data or preexisting maps and enables the autonomous system to follow a reference route - be it recorded from an earlier drive or from another vehicle currently leading the way. The system for route following consists of two components: the first component takes care of the local obstacle avoidance and path planning by computing a multi-layer grid-based representation of the environment, which is given as input to a trajectory planning module. Data is accumulated in two-dimensional Cartesian grids over time while the vehicle moves, giving a more complete and denser environment representation than what can be achieved from a single time step alone. For an increased field of view, multiple stereo camera systems are used. Based on a stereo camera sensor model, each stereo system computes a local measurement grid which is then fused into the accumulated grid. The grid contains obstacle probabilities as the main feature for planning collision-free paths. Additional layers, such as the elevation, terrain slopes, and color information are computed for planning a more comfortable, robust path through the local environment. Due to the large amount of data from the stereo cameras, the processing must be efficient in order to allow real-time use. The second component is the landmark-based mapping and localization system, enabling the route following behavior in unknown environments. Without the need for global positioning data, it uses both low-level image features and object landmarks to create a map and localize on it. In a vehicle-following scenario, the leader transmits the landmarks together with the driven route in an odometry frame to the following vehicle using vehicle-to-vehicle (V2V) communication. The follower uses the available data to build a map representation. By comparing the map with the landmarks perceived itself, it can localize and therefore follow on the leader’s driven route. In this thesis two different landmark types are developed that are extracted from camera images and used for mapping and localization. Experimental results show the applicability of the proposed system – both for local obstacle avoidance using the multi-layer grid and for landmark-based localization on a reference route.
«Autonomous vehicles often rely on highly accurate map data for navigation purposes. However, such maps are of limited availability and restrict the vehicle’s region of operation to already mapped areas. This thesis describes a vision-only perception system for autonomous obstacle avoidance and route following in unknown environments. It works without the need for global positioning data or preexisting maps and enables the autonomous system to follow a reference route - be it recorded from an ear...
»