Object Detection and Semantic Segmentation for a Small Low-cost Vehicle

This contribution presents a demonstrator vehicle for automated driving in different or even unknown and rough terrains. Covering different terrains is an unique feature which separates this approach from the classic automotive industry which mainly focuses on driving on existing roads. This makes our approach highly interesting for commercial vehicles which are not limited to roads as well as those which have to drive on roads without reliable maps such as after catastrophes or simply in non-cartographed areas like construction sites or production plants. The main focus here is the perception. Thus, the object detection in 3D and the semantic segmentation will be covered. The former is needed to not only detect obstacles and their location but also whether to avoid or to run over them. The latter gives information about the surrounding: E.g. which areas of the ground are driveable. This is important if not only the road is considered as viable driving surface. To leave the level of an only academic feasibility, common commercial off-the-shelf (COTS) hardware is used and attention is paid on cost-benefit ratio. For the implementation, different neural networks are compared and finally used.