An End-to-End ConvLSTM-based Method for Point Cloud Streaming Compression

The processing and transmission of point cloud frame sequences is an important part of the applications of 3D LiDAR. However, due to the disorderliness and irregularity of the huge amount of point cloud data collected by 3D LiDAR sensor, finding an effective method to compress the point cloud data to a small volume is an urgent problem. In this paper, we propose spatio-temporal features sensitive neural network SPCCNet with an encoder-decoder structure to compress point cloud streams. To reduce information loss in point cloud preprocessing, we propose a unique convolution method on point sets. The curvature and density information are introduced to SPCCNet to enhance the raw point cloud data. Besides, the designed ConvLSTMlm and a Squeeze-and-Excitation (SE) Block are embedded to help SPCCNet learn the effective features of point cloud sequences. Experimental results show that compared with other methods, our SPCCNet can compress point cloud data with a higher compression ratio at an acceptable noise level.

Recommended citation: J. Long, M. Feng, B. Li, Y. Ling, C. Wu, K. Huang, and M. Cui*, An End-to-End ConvLSTM-based Method for Point Cloud Streaming Compression, 2024 IEEE International Conference on Advanced Robotics and Mechatronics (ICARM).
Download Paper | Download Slides