We present an odometry approach relying on monocular panoramic video over several kilometers in urban streets with midday traffic. While current approaches rely on pose updates from reconstructed proximal landmarks, we introduce a bio-inspired approach where distal landmarks provide robust estimates of absolute orientation obtained from 2D-2D correspondences. Our approach is frame incremental and does not contain any iterative or batch steps.
Mapping requires solution of the loop closing problem. Loop closing as well as arbitrary post-mapping localization are place recognition capabilities. Starting from a pure appearance based approach for place recognition, we investigate geometric and structural methods for matching images based on location. Results are shown in km-long trajectories in mid-day Philadelphia traffic.