PVENet: Point Voxel Encoder Network for Real-Time Classification of Lidar Point Cloud Segments

For a vehicle equipped with an automated driving system, object classification is an important feature for safe operation. Several lidar based object detection approaches use heuristic methods to extract a point cloud segment for each object which has to be classified. In this paper, we show a new method for the classification task using a CNN on voxels, while the features of each voxel are calculated using an encoder network, which efficiently processes point clouds. This results in an architecture with small voxel grid sizes (8 × 8 × 8) at very low computational effort for real-time point cloud classification. Our classifier provides an accuracy of 96.8% on the Stanford Track Collection. Furthermore, the processing time of our classifier consumes only 0.43 milliseconds per point cloud segment, which makes this method applicable for real-time classification tasks.