3D Point Cloud Enhancement using Unsupervised Anomaly Detection

3D point cloud is increasingly getting attention for perceiving 3D environment which is needed in many emerging applications. This data structure is challenging due to its characteristics and the limitation of the acquisition step which adds a considerable amount of noise. Therefore, enhancing 3D point clouds is a very crucial and critical step. In this paper, we investigate two promising unsupervised techniques which are One-Class SVM (OCSVM) and Isolation Forest (IF). These two techniques optimize the separation between relevant/normal points and irrelevant/noisy points. For evaluation, three metrics are computed, which are the processing time, the number of detected noisy points, and Peak Signal-to-Noise (PSNR) in order to compare the both proposed techniques with one of the recommended filters in the literature which is Moving Least Square (MLS) filter. The obtained results reveal promising capability in terms of effectiveness. However, OCSVM technique suffers from high computational time; therefore, its efficiency is enhanced using modern Graphics Processing Unit (GPU) with an average rate improvement of 1.8.