Benchmarking lossy octree decimation and predictive quantization for cooperative perception at 20 Hz over 5G NR PC5.
A standard automotive solid-state LiDAR generates between 1.5 and 3 million points per second, yielding raw data streams exceeding 250 Mbps. Direct C-V2X PC5 channels, operating within 10 to 40 MHz allocated ITS spectrum, cannot sustain raw uncompressed point cloud streaming across congested vehicular environments.
V2Sensor applies hierarchical spatial octree decomposition paired with predictive velocity-vector delta encoding. Points falling onto static asphalt or buildings are pruned using pre-existing HD map priors, streaming only non-ground dynamic obstacles and unknown surfaces. ``` +-----------------------+ Octree Compression +-----------------------+ | 32-Beam Automotive | --------------------------------> | 5G NR Sidelink Packet | | LiDAR (250 Mbps) | >95% Data Reduction | (<10 Mbps Stream) | +-----------------------+ +-----------------------+ | v Remote Vehicle / RSU Reconstructed Point Cloud ```
Extensive vehicle-to-vehicle testing demonstrated that compressed 100 KB/s point cloud streams achieved 98.4% 3D object detection accuracy (mAP) compared to uncompressed raw LiDAR streams, while fitting within standard sidelink resource allocations.