Abstract
Abstract. Real world localization tasks based on LiDAR usually face a high proportion of outliers arising from erroneous measurements and changing environments. However, applications such as autonomous driving require a high integrity in all of their components, including localization. Standard localization approaches are often based on (recursive) least squares estimation, for example, using Kalman filters. Since least squares minimization shows a strong susceptibility to outliers, it is not robust.In this paper, we focus on high integrity vehicle localization and investigate a maximum consensus localization strategy. For our work, we use 2975 epochs from a Velodyne VLP-16 scanner (representing the vehicle scan data), and map data obtained using a Riegl VMX-250 mobile mapping system. We investigate the effects of varying scene geometry on the maximum consensus result by exhaustively computing the consensus values for the entire search space. We analyze the deviations in position and heading for a circular course in a downtown area by comparing the estimation results to a reference trajectory, and show the robustness of the maximum consensus localization.
Highlights
Tackling localization tasks including the fusion of different sensor measurements using Kalman filters is a common approach (Chen, 2012)
Apart from considering the highest consensus set in the accumulator, we want to use the gained entire knowledge about the search space to evaluate the quality of the localization result
We have shown that using maximum consensus based on LiDAR provides great potential for high integrity localization
Summary
Tackling localization tasks including the fusion of different sensor measurements using Kalman filters is a common approach (Chen, 2012). Heuristic approaches, based on a nonsmooth penalization and the alternating direction method of multipliers, respectively, have recently been reported by (Le et al, 2017) While they show good results and attractive execution times, especially at higher outlier rates, they do neither ensure global optimality nor being within certain bounds from the optimal result. (Chen et al, 1999) and (Aiger et al, 2008) use randomized heuristics, (Li and Hartley, 2007) and (Yang et al, 2013) are using maximum likelihood frameworks, and (Chin et al, 2014) and (Parra Bustos et al, 2016) are using a geometric matching criterion, striving for globally optimal point cloud registration using BnB. In this paper we show (i) a localization based on the registration of two point clouds, a sparse, ‘car sensor’ cloud from a Velodyne VLP-16 scanner, and a dense, high resolution ‘map’ point cloud from a mobile mapping campaign, (ii) a high integrity localization method using the maximum consensus criterion, leading to a globally (within search radius) optimal solution, and (iii) an assessment of the effects of scene geometry on the distribution of the maximum consensus score
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
