SC-IOCI: Multiple Infrastructure LiDAR-Based 3-D Object Detection via Intermediate Output Compression and Integration in Split Computing

Abstract

3D object detection based on LiDAR point cloud data and deep neural networks (DNNs) plays a critical role in autonomous driving systems. Although state-of-the-art models achieve high accuracy, deploying them on edge devices remains challenging due to their computational complexity and latency. Furthermore, single-LiDAR-based object detection suffers from occlusion and limited viewpoints. Infrastructure LiDARs can address these limitations; however, efficiently utilizing multiple sensors under computational and communication constraints remains challenging. This paper proposes SC-IOCI, a multiple infrastructure LiDAR-based 3D object detection method on edge devices using Split Computing with intermediate output compression and integration. Each edge device processes local point cloud data using early DNN layers to extract intermediate features, which are compressed and transmitted to a central edge server, so that the raw point clouds remain on the edge devices that acquired them. The server decodes, aligns, and integrates the received intermediate outputs to perform inference. This design reduces edge-device workload, shortens inference latency, and enables a flexible trade-off between detection accuracy and communication efficiency. Experiments demonstrate that integrating intermediate outputs from multiple infrastructure LiDARs improves detection accuracy, particularly at higher Intersection over Union (IoU) thresholds. In addition, the proposed method achieves up to 3.4 times faster inference and reduces communication data size by approximately 75% while preserving accuracy.