Effective software security enhancement using an improved PointNet++

As common three-dimensional (3D) data, point clouds have wide application prospects in many fields. The point cloud disorder problem is solved by the PointNet neural network using a symmetric function, which insensitivity to the input order enables PointNet to process the original point cloud data directly. The ability to extract local features was enhanced by introducing the PointNet++, furnishing a better solving capacity of 3D vision problems and improved intelligent driving software security. This paper analyzes the PointNet++ implementation principles and improves its local feature extraction capability by the proposed density-related farthest point sampling (DR-FPS) algorithm, mitigating some limitations of the conventional FPS algorithm so that the sampling results can better express the feature information of point cloud data. The accuracy of the proposed DR-FPS algorithm applied to five-category and ten-category classification tasks exceeded that of the conventional FPS by 4.4 and 5.6%, respectively. Finally, a positive correlation between the accuracy increment and the number of classification categories was revealed. The results are instrumental to intelligent driving software security enhancement.