Current vision-based localization approaches enable reliable positioning in areas where global navigation satellite systems (GNSS) fail due to multipath and shadowing effects. These approaches require an up-to-date map. It seems promising to update such maps iteratively after passing the mapped area again. However, bundling more and more passes into the existing map leads to unbounded computation and memory complexity. Herein we propose an iterative optimization approach to create highly accurate maps comprising any number of drives with constant computation complexity. The optimization bases on keypoint correspondences matched between the recorded images from multiple drives. First, each new drive is reconstructed separately by a sliding window bundle-adjustment. Thereafter, the estimated trajectory is divided into disjoint clusters. To align the new drive to the current map, we optimize pairs of clusters which are interconnected through loop-closure or inter-drive correspondences. We derive pose differences from all clusters to estimate the final map poses. For global accuracy, we add GNSS measurements from a low cost receiver. We show in our experiments that the approach enables a joint estimate of the trajectories and landmarks from numerous city-scaled passes within several hours on desktop computers.
Efficient Multi-Drive Map Optimization towards Life-long Localization using Surround View
2018-11-01
2364288 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch