Fast and Accurate Ground Plane Detection for the Visually Impaired from 3D Organized Point Clouds

Fast and Accurate Ground Plane Detection for the Visually Impaired from 3D Organized Point Clouds

Summary

This paper proposes an algorithm to help visually impaired people navigate in a fast, safe, reliable and independent manner. Using RGB-D scanners, an enhanced RANSAC algorithm is proposed to eliminate common RANSAC problems. The proposed algorithm is able to detect the ground and obstacles facing the visually impaired. The algorithm consists of three main stages: data preprocessing, ground segmentation and object detection.

Specific steps

1. Data preprocessing
   Straight-through filter: Abandon distant points (the distant points are noisy)
   Voxelization: Divide the three-dimensional space point cloud into 3 3 3 grid blocks, and each block is represented by the centroid of the points in the block,
2. Ground Segmentation
   Normal Estimation: Estimated from a point in a neighborhood of surrounding points (also called k-neighborhood).
   Plane segmentation using RANSAC and surface normals
3. Object Detection and Clustering
   First, the detected ground is removed and the remaining objects are treated as obstacles. Apply Euclidean segmentation (using pcl::EuclideanClusterExtraction class) to the remaining points to cluster the objects. Check the distance between two points and merge them if the distance is below a threshold.