Depthwise Over-Parameterized CNN for Voxel Human Pose Classification

Abstract

Light Detection and Ranging (LiDAR) capture objects and backgrounds using a laser sensor, producing unstructured points in 3-dimensional called point clouds (PC). However, captured human pose PC is limited partially due to the LiDAR scan. The only information in the scanned area exists. Due to the inadequacy of PC data, it is challenging to classify such data. In this paper, we proposed a solution to overcome those problems. It is a novel depthwise over-parameterized (DOConv) embedded into a simple CNN. The raw PCs are converted into a 3D voxel in the input layer. In the convolutional (Conv) layer, the regular Conv is substituted with a-three layered DOConv. Lastly, to assess the performance of our model, we commence an evaluation with multiple classifier algorithms in ModelNet40 and our human pose dataset. Accuracy, loss, recall, precision, F1-scores, and Geometric mean are engaged as performance indicators. To sum up, our model outperformed all compared classifiers in accuracy for the primary dataset by 87.06 % and ModelNet40 by 68.68%.