Cluster-based Object Detection System with Scalable Performance for Autonomous Driving
Hongsuk Kim, Yongseong Lee, Jangho Shin, Jong-Chan Kim · 2023
Due to the unprecedented computing requirement of autonomous driving applications, in-vehicle computing architecture is going under tremendous changes. For that, emerging computing units such as graphics processing units (GPUs) and neural processing units (NPUs) are employed, trying to satisfy such computing requirement. However, since the computing power of a single node is inherently limited, we cannot scale up beyond a single node's computing capacity. To make a scalable computing architecture for computing-hungry applications, we propose to employ the cluster-based computing architecture that has been used in server-side applications for decades. As an initial effort, we develop a prototype object detection system by clustering three compute nodes equipped with an Nvidia SoC with an integrated GPU, connected through an ethernet bus and a CAN bus. A GigE Vision camera is connected to the ethernet bus and multicasts images to the three nodes. For each node to select its assigned images in a round-robin manner, a distributed consensus scheme is proposed. Our prototype implementation demonstrates a linear scalability of its frame rate up to three nodes with no additional delay overhead caused by the clustering architecture. By that, this study shows the potential of the cluster system for autonomous driving applications.