Utilization of Quadnocular Stereo Vision for Simultaneous Localization and Mapping in Autonomous Vehicles

The introduction of autonomous vehicles in commercial applications has an enormous potential to drastically improve both working conditions and productivity. Among the many challenges of realizing autonomous driving, precisely determining the vehicle's pose and constructing an accurate map of an unknown environment might be the two most crucial to overcome. Simultaneously dealing with both these problems is known as SLAM, simultaneous localization and mapping. We have designed and implemented a stereo camera system with which SLAM is performed using only visual information. The proposed system achieves multi-range depth resolution by utilizing four cameras and is implemented exclusively with the help of open-source libraries. Depth images from separate camera pairs are computed in real time with a correlation based block matching algorithm and merged employing a perspective transform. The joint depth map and the corresponding RGB image are then used by a graph-based SLAM algorithm to produce a trajectory estimate and a probabilistic 3D voxel grid map. The system was furthermore evaluated on simulated and real data. The evaluations suggested that with the current configuration it is not motivated to merge depth images from more than two camera pairs, since adding an extra camera pair considerably decreases the frame rate without necessarily improving the quality of the resulting depth map. A number of different image feature extractors were investigated which revealed that binary feature descriptors are suitable for landmark representation in visual SLAM. We can conclude that the system is indeed a fully working prototype, but in order to increase robustness and accuracy to a degree that it is useful in a commercial application a number of future improvements are needed