Point Cloud Subsampling Parallelization for Unified Memory Platforms

TL;DR

This paper introduces a point cloud decimation method optimized for unified memory platforms, enabling efficient data reduction for robotic environment mapping on resource-limited embedded systems.

Contribution

It presents a novel point cloud decimation implementation that leverages unified memory architectures for improved performance on embedded systems.

Findings

Unified memory architectures enhance data communication efficiency.

The method performs well across different grid sizes and scenarios.

Embedded systems benefit from shared memory for point cloud processing.

Abstract

The exploration of unknown environments using robots is a task that integrates different areas such as location, mapping, and planning. For mapping, there are numerous methods to represent the environments through which a robot can travel, in two and three dimensions. The probabilistic occupation grid, Octomap, and STVL can be mentioned among the most important in recent years. Nowadays, RGB-D cameras are widely used to generate a detailed representation of the environment. RGB-D camera measurements present a large volume of data, which must be reduced in order to be used in platforms with limited computing resources. This work presents an implementation of the point cloud decimation method capable of being executed on platforms with unified memory. It consists of reducing the point cloud iteratively using a subdivision of space. Results were obtained for different sizes of grids, platforms, and scenarios, both real and simulated. The results indicate that in embedded systems it is convenient to have architectures that share memory between CPU and GPU to optimize data block communication processes.