Probabilistic voxelated three dimensional grid map for simultaneous localization and mapping

Abstract

With the increase of robot chassis mobility and the abundant solution for simultaneous localization and map-building, an autonomous robot has unprecedented opportunity to explore an environment in situ. However, the mobility of the robot is hindered by the limited availability of three dimensional probabilistic model of the environment. In so doing, this research proposed and investigated the potential of using probabilistic voxelated three dimensional grid map that can produce three dimensional map probabilistically by incorporating stochastic nature of sensor reading and robot movements. The map is modelled to give direct probability of an occupied space. The grid cells are voxelized and embedded with relax logit function to emanate probability value of an occupied space. The performance of the probabilistic voxelated three dimensional grid map was tested by using scans collected a priori. These scans act as a kernel to the registration technique. Two separate sets of map were reconstructed using pose estimate obtained from Bayesian filters. The result was compared between maps produced under recursive Bayesian filter and Bayes' rule filtering with referenced to the geometric information of the controlled environment. The probabilistic voxelated three dimensional grid map restored the geometric information of the environment with accuracy of 0.87. The runtime of the method converged to 0.23 s after the fifth map registration. The method also compressed the metrical data from Hokuyo UTM-30LX by two decades in logarithmic scale. By using continuous probability value to represent occlusion, the map has the potential to be used with any Bayesian filtering family. The potential of the probabilistic voxelated grid map to be used as path planning method was also observed since the probability value in each grid cell of the map behaves similarly to vector field.