Hybrid Point Cloud Semantic Compression for Automotive Sensors: A Performance Evaluation

Abstract

In a fully autonomous driving framework, where vehicles operate without human intervention, information sharing plays a fundamental role. In this context, new network solutions have to be designed to handle the large volumes of data generated by the rich sensor suite of the cars in a reliable and efficient way. Among all the possible sensors, Light Detection and Ranging (LiDAR) can produce an accurate 3D point cloud representation of the surrounding environment, which in turn generates high data rates. For this reason, efficient point cloud compression is paramount to alleviate the burden of data transmission over bandwidth-constrained channels and to facilitate real-time communications. In this paper, we propose a pipeline to efficiently compress LiDAR observations in an automotive scenario. First, we leverage the capabilities of RangeNet++, a Deep Neural Network (DNN) used to semantically infer point labels, to reduce the channel load by selecting the most valuable environmental data to be disseminated. Second, we compress the selected points using Draco, a 3D compression algorithm which is able to obtain compression up to the quantization error. Our experiments, validated on the Semantic KITTI dataset, demonstrate that it is possible to compress and send the information at the frame rate of the LiDAR, thus achieving real-time performance.

CACTUS: Content-Aware Compression and Transmission Using Semantics for Automotive LiDAR Data

“…Codec parameters for can be manually tuned (e.g., the Quantization Parameter (QP)) in order to vary the reconstruction accuracy and the allocated bitrate. This larger amount of degrees of freedom makes the approach more flexible with respect to [ 17 ], where three fixed configurations are adopted and only a limited set of information is transmitted. In this way, the user or the algorithm can choose to lower the quality of semantic classes (or even completely remove them) that are less useful in the considered scenario (e.g., a self-driving car might not need to precisely know the position of the vegetation outside the road).…”

“…In the work by Varischio et al [ 17 ], Draco [ 13 ] is used in conjunction with Rangenet++ [ 55 ] to perform the encoding. In particular, the authors define three compression levels where parts of the PC are progressively removed to allow to perform compression in real time.…”

“…As an example, the work by Huang et al [ 14 ] uses an entropy model to exploit the remaining redundancy after an octree decomposition of voxel volume [ 15 ]; Tu et al [ 16 ] use Recurrent Neural Networks to progressively encode residuals, while in the work by Varischio et.al. [ 17 ], the most meaningful parts of the PC are transmitted after a segmentation of the input PC. Other meaningful approaches that are not specifically targeted toward LiDAR but that are worth mentioning are deep learning (DL)-based Point Cloud Coding (PCC) approaches [ 18 , 19 , 20 , 21 , 22 , 23 , 24 ] that generally exploit convolutional autoencoders for efficient features extraction and decoding.…”

“…Most of the works that were explored in the literature address the compression task in a very general way and thus do not exploit semantic data for transmission, since it is not always available. To the best of our knowledge, up to the publication of this paper, the only work that makes use of this additional information is [ 17 ]. However, the authors fail to fully exploit it since, in the paper, semantics only serve the purpose of selecting pieces of the PC that should be discarded in order to obtain lower compression rates and faster decoding.…”

Cloud-based XR Services: A Survey on Relevant Challenges and Enabling Technologies

“…Draco aims to improve the storage and transmission of 3D graphics by compressing meshes and point-cloud data. The huge impact of Draco is depicted in current bibliography, as it influences and drives the design of alternative compression/decompression techniques [52], [53], [54], [55], [56], [57], [58], [59], [60], [61]. The framework uses a kd-tree to efficiently store data corresponding to points, connectivity information, texture coordinates, color information, normals and any other generic attributes associated with geometry.…”