3D object classification using 2D perspectives of point clouds

Abstract

As three-dimensional point clouds become an increasingly popular for representing spatial data, with it comes a need to classify objects represented in this form. This paper proposes an image feature representation technique which classifies 3D point clouds by using several 2D perspectives, and then using YOLO (v3) as object classification tool to classify the rasters. Existing works have had limited success using this technique with only one perspective, likely due to the considerable information loss during the projection. We hypothesize that multiple projections mitigate these effects by capturing lost data in other perspective projections. Our method is effective in classifying pedestrians and vehicles in the Sydney Urban Objects data set, where our technique has achieved a classification accuracy of 98.42 % and an f1 score of 0.9843, which are considerably higher than existing methods.