Fuzzy-Aided Compression: An Efficient Point Cloud Compression Algorithm for Collaborative 3D Object Detection of Autonomous Driving

Current collaborative 3D object detection based on light detection and ranging (LiDAR) sensors face challenges when processing the substantial volume of LiDAR points. They typically adopt one of two approaches: aggregating all points and randomly discarding a specific number of points, or relying on deep neural networks to extract feature embeddings from each agent and then performing fusion. The former approach cannot guarantee optimal point selection, while the latter incurs substantial computational costs, making it impractical for real-time autonomous driving systems. To tackle these challenges, we introduce Fuzzy-aided Compression (FAC), an efficient yet effective LiDAR points compression algorithm for collaborative 3D object detection of autonomous driving. FAC integrates a fuzzy logic-based point selector and a distance-aware convolution plugin. In particular, FAC starts with a fuzzy logic module for point selection, using Takagi-Sugeno-Kang to assess point importance. To counter point reduction’s information loss, we add distance heatmaps to convolution layers for better sensor data integration. This setup is merged into a unified, trainable framework linking point selection with object detection. Extensive experiments and evaluations were conducted on OPV2V and NuScenes datasets. Empirical results indicate that FAC surpasses state-of-the-art baselines, achieving a higher mean average precision for detection while using fewer LiDAR points and demanding less computational time.