Point cloud object recognition method via histograms of dual deviation angle feature

Abstract

ABSTRACT LiDAR point cloud object recognition is widely adopted in 3D perception applications and plays an important role in remote sensing, autonomous driving, robotics and other fields. We propose a novel point cloud object recognition method based on histograms of dual deviation angle feature, which we call the HDDAF. First, we use the normal alignment radial feature to sample feature points. Second, we construct the local reference frame (LRF) for the point cloud. Third, we calculate the connecting line between the barycentre and LRF origin, and calculate the deviation angle between connecting line and the LRF and deviation angle between connecting line and point cloud normal vectors. Fourth, we construct feature histograms of the two kinds of deviation angle. Finally, we construct a model library using the HDDAF algorithm to transform the point cloud data into a feature histogram, and we transform the point cloud of unknown objects into a feature histogram in the same way. We match the histogram of unknown objects with the model library histogram to recognize the unknown object. We rigorously test the proposed method on B3R and ModelNet40 standard datasets and real scenes, and compare with those of 10 other methods. For standard, noise and real scene data, the precisionrecall curve of the proposed method is superior to that of the comparison methods. The average computation times of the proposed method on the two datasets and real scenes are 0.221, 0.051 and 0.448 seconds, which are shorter than those of the comparison methods. The results show that the proposed method is highly descriptive, robust and computationally efficient.