Core Concepts
A computationally efficient pipeline for detecting obstacles in point cloud data using a LiDAR sensor on a resource-limited robot platform.
Abstract
The paper presents a pipeline for efficient obstacle detection in point cloud data obtained from a LiDAR sensor, designed to run on a computationally constrained robot platform.
The key steps of the pipeline are:
Preprocessing the point cloud data:
Outlier removal using Voxel Grid filtering
Floor removal using RANSAC
Filtering based on an occupancy map to keep new observations up-to-date
Clustering the preprocessed point cloud to group points into obstacle blocks.
Determining the optimal rotation of the blocks based on observations to avoid occupying unobserved space.
The pipeline was implemented and tested on a Raspberry Pi-based robot platform running ROS2. Qualitative evaluation showed the pipeline's ability to detect obstacles in real-time, with performance comparisons to 3D object detectors running on GPUs.
Future work will focus on extending the pipeline to include obstacle classification, allowing the robot to respond differently to different types of obstacles (e.g., moving humans vs. static shelves).
Stats
The pipeline was able to process point cloud data at 11ms per frame on the Raspberry Pi CPU, compared to 40ms and 171ms for two 3D object detectors running on GPUs.
Quotes
"The pipeline allows to find obstacles in point cloud."
"The pipeline is suitable for obstacle search in real-time."