PARTICLE FILTER BASED DETECTION OF DYNAMIC OBJECTS ON AN ACCUMULATED OCCUPANCY MAP

Abstract

The paper considers the problem of detecting dynamic obstacles on the accumulated occupancy map generated by the computer vision system of a mobile robot. The purpose of this research is to improve the quality of the obstacle detection algorithm by adding a particle filter to find moving objects from the map data. In the paper, the problem of correct accumulation of data in the occupancy map and reducing the delay in updating the map cells in which the object moves is solved. The modification of the particle filter presented in the paper is able to work correctly with dynamic obstacles in a wide range of speeds; it is resistant to outliers caused by random generation of the initial particles velocities, and is workable under real conditions in real time in an environment with a lot of moving objects. A heuristic has been created that reduces the number of misclassifications in occluded areas. It is shown that the algorithm for detecting dynamic objects in the map is invariant to the type of sensors used in the vision system, and an implementation combined with an accumulated occupancy map is described. The algorithm is implemented and tested on board an autonomous mobile robot, as well as on an open dataset. The article also provides a comparison with other approaches of dynamic obstacles detection, as well as calculated performance metrics for all analyzed methods for computers based on the GPU Nvidia RTX 3070 and Jetson AGX Xavier. Promising directions for further research to improve the presented algorithm are formulated.