Probabilistic inverse sensor model based Digital Elevation Map creation for an omnidirectional stereovision system

Abstract

The objective of the paper is to present an original solution for building a high accuracy Digital Elevation Map (DEM), from down-looking omnidirectional stereo system data, used for surrounding perception. For this reason, an accurate probabilistic inverse sensor model of the omnidirectional stereo sensor is estimated based on training data. The obtained model considers not only the Gaussian spread of 3D points but also the systematic translations and errors from the calibration and rectification processes. The inverse sensor model is obtained by calculating the prior probabilities of 3D points corresponding to each DEM cell and the direct sensor model, describing the way measurements are acquired. The direct sensor model is calculated using an umbrella based modified Shepard trilinear interpolation of individual measurements in space. The results of the interpolation (σx, σy, σz, μx, μy, μz,) are stored in a 3D lookup table which performs a discretization of 3D space into cuboids. For each 3D point the probability of correspondence to the neighboring cells is calculated and the obtained values are added to the height histogram of each cell. Instead of adding to a single bucket in the histogram, the contribution is spread based on the standard deviation of the height. In order to increase the contribution of individual points in sparse areas and to decrease it in dense areas, the relative density of 3D points in local patches is precomputed and is used as a decreasing exponential term. Based on the obtained models, an improved DEM creation algorithm is applied. The obtained elevation map provides better results both in terms of accuracy and detection rate.