METHOD FOR GENERATING A MOBILE ROBOT OCCUPANCY MAP FROM MULTISPECTRAL VISION SYSTEM DATA

Abstract

This paper addresses the problem of generating occupancy maps of the surrounding space for robotic
platforms using data from a multispectral vision system. The aim of the work is to qualitatively improve
the generated passability model by combining geometric and semantic data from lidars and stereo
cameras, as well as direct velocity measurements from millimeter-wave radars. The presented algorithms
and their modifications are universal to the data source and do not require physical synchronization of
sensors. The paper solves the problem of constructing both static a priori and real-time dynamic occupancy
maps. An approach for combining an a priori semantic map with the one generated in the runtime is
proposed. Approaches for accumulating and updating semantic information in the maps are described.
The problem of detecting dynamic obstacles in occupancy maps based on a modified particle filtering
algorithm is also considered. The combined method described in the paper increases the accuracy of dynamic
obstacle detection and enables correct obstacle detection even if the dynamics detection algorithm
fails. Metrics for quantifying occupancy maps are defined. The developed algorithm has been tested on
Semantic KITTI, nuScenes open datasets in the automotive data domain, and on a small service cleaning
robot both in the CARLA simulator and in real-world conditions with active pedestrian traffic. The software
implementation of the algorithm runs in real time on Jetson AGX Xavier and Jetson AGX Orin embedded
computers.