PointCompress3D - A Point Cloud Compression Framework for Roadside LiDARs in Intelligent Transportation Systems

Walter Zimmer^*, Ramandika Pranamulia^*, Xingcheng Zhou, Mingyu Liu, Alois C. Knoll

Technical University of Munich (TUM), Germany
^*Indicates Equal Contribution

Visualization of our point cloud compression and streaming pipeline, showcasing raw LiDAR point clouds on the left, progressing through preprocessing, compression, storage, training, and 3D object detection, merging in the streaming of compressed point clouds with 3D detection results.

Overview

PointCompress3D is the first point cloud compression framework for roadside LiDARs. With this framework you can select a point cloud compression algorithm to store roadside point clouds efficiently and stream them in real-time using the state-of-the-art loosy and lossless compression algorithms.

In summary:

We propose a point cloud compression framework for roadside infrastructure LiDAR sensors and a dev-kit.
We provide an in-depth comparison of state-of-the-art compression methods on the SemanticKITTI, Ford and TUMTraf dataset family.
We extend existing compression methods to make them compatible with our roadside Ouster LiDAR sensors
We perform extensive experiments and ablation studies on our real TUMTraf Intersection and TUMTraf V2X Coop. Perception dataset and evaluate six metrics.
We open-source our framework, which contains the point cloud projection and compression module and provide a project website with video results.

Abstract

Efficient management of point cloud data is mandatory within the field of intelligent transport systems. Requirements to efficiently store, stream, and do object detection towards the point cloud data are demanded in a real-time manner. To address this challenge, three state-of-the-art compression methods are explored and evaluated using our real-world-based TUMTraf dataset family. We achieve a frame rate of 10 FPS while maintaining compression size below 250 Kb and keeping the performance of the object detection par with the original data. In this work, we present a comprehensive assessment of each method, outlining their efficacy and suitability for practical implementation. Future work includes the deployment on the live system. Finally, we provide code, weights, and visualization results on our website: https://pointcompress3d.github.io.

Qualitative Results

Qualitative results on our TUMTraf Intersection dataset. From left to right: a) Side view of a truck and a car in the original point cloud. b) Experiment E1: Compressed point cloud with a subsampling distance of 3.0 and a minimum kernel radius of 1.5. c) Experiment E3: Compressed point cloud with a subsampling distance of 1.0 and a minimum kernel radius of 1.2. Both compressed images are generated with max. 30,000 points and a grid size of [40, 40, 15] m.

Left: Stream of original point clouds of the TUMTraf A9 Highway dataset. Right: Compressed stream of point clouds.

Ablation Studies

Adjusting the minimum kernel radius parameters by 0.5 times and 2 times the original value, respectively.

We set the kernel radius in encoding block to 10 and 0.05 (top row) while keeping the decoding block values constant at 0.05. Subsequently, we set kernel radius in the decoding block to 1 and 0.01 (bottom row) while keeping the encoding block values constant at 1.0.

Visualization of ablation study for the max. number of points. The point cloud shows a side view of a truck and a car with 5,000, 12,500, 25,000, 50,000, 100,000, and 200,000 points.

Left: Stream of original point clouds of the TUMTraf Intersection dataset. Right: Compressed stream of point clouds.

Quantitative Results

Grid Size	Max. Points	Enc. (ms)	Dec. (ms)	Enc. VRAM (GB)	Dec. VRAM (GB)	PSNR d1	PSNR d2	BPP	mAP 3D
8x8x3	50,000	220	2.7	3.9	3.9	15.32	23.68	7.48	13.32
	100,000	410	2.7	6.0	4.3	21.81	30.13	11.72	17.29
	200,000	520	2.8	5.5	7.1	24.18	32.88	13.57	19.39

16x16x6	50,000	230	2.6	3.0	2.7	-7.81	-0.19	5.17	12.21
	100,000	350	2.5	4.3	3.7	-3.59	3.87	7.03	19.50
	200,000	510	2.7	6.8	4.0	-1.46	6.31	7.73	20.91

24x24x9	50,000	240	2.5	2.2	2.7	-5.57	0.99	3.90	13.59
	100,000	410	2.6	5.5	3.1	-1.87	5.50	4.94	19.75
	200,000	530	2.7	6.6	3.2	-0.18	7.47	5.31	19.32

Parameter tuning for DEPOCO on the max. number of points and the grid size. We use the TUMTraf Intersection dataset to find the best parameters and evaluate the compression method on the TUMTraf V2X Cooperative Perception dataset.