An Auto-Encoder-Based Scheme for LiDAR Point Cloud Compression

Light Detection and Ranging (LiDAR) has been widely used in various fields in recent years, and the resulting large amount of point cloud data storage has become an issue to be solved. To solve this problem, the commonly used point cloud coordinate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(x,\ y,\ z)$</tex> is transformed to the spherical coordinate <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(r,\ \omega,\ \alpha)$</tex> in this paper. We then construct the so-called two-dimensional slices, which constitutes the fundamental coding unit of the point cloud, along azimuth direction with certain time interval. After that an auto-encoder is applied for the prediction of the slice to be compressed. The generated feature from the auto-encoder is further encoded with a sparse coding technique, and then entropy encoded with a run-length coding scheme with Golomb code. The prediction residues caused by the auto-encoder are further encoded with an error compensator to get smaller prediction residues. To evaluate the performance of the proposed approach, the Kitti dataset, a commonly used benchmark for 3D point cloud, is applied. Experimental results show that an average reconstruction error of 0.4 cm with 7.6 bpp (bits per point) can be obtained. A very good compression performance in terms of bit rates can be achieved when low reconstruction error is required.