3-D Point Cloud Object Detection Based on Supervoxel Neighborhood with Hough Forest Framework

Hanyun Wang,Cheng Wang,Huan Luo,Peng Li,Yiping Chen,Jonathan Li

DOI: https://doi.org/10.1109/jstars.2015.2394803

IF: 4.715

2015-01-01

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

Abstract:Object detection in three-dimensional (3-D) laser scanning point clouds of complex urban environment is a challenging problem. Existing methods are limited by their robustness to complex situations such as occlusion, overlap, and rotation or by their computational efficiency. This paper proposes a high computationally efficient method integrating supervoxel with Hough forest framework for detecting objects from 3-D laser scanning point clouds. First, a point cloud is over-segmented into spatially consistent supervoxels. Each supervoxel together with its first-order neighborhood is grouped into one local patch. All the local patches are described by both structure and reflectance features, and then used in the training stage for learning a random forest classifier as well as the detection stage to vote for the possible location of the object center. Second, local reference frame and circular voting strategies are introduced to achieve the invariance to the azimuth rotation of objects. Finally, objects are detected at the peak points in 3-D Hough voting space. The performance of our proposed method is evaluated on real-world point cloud data collected by the up-to-date mobile laser scanning system. Experimental results demonstrate that our proposed method outperforms state-of-the-art 3-D object detection methods with high computational efficiency.