Reducing Communication Cost and Latency in Autonomous Vehicles with Subscriber-centric Selective Data Distribution

Driving automation has become a major cost factor in automotive design. High computation demand for machine learning (ML) applications and growing sensor resolution and data rate require expensive hardware technology and networking. Newer network technologies and topologies can at best compensate the growing communication demand, but the network still accounts for a complex wiring harness with many dedicated sensor cables. In this paper, we exploit the context specific sensor data access of ML perception applications to minimize the sensor data traffic. For that purpose, we extend the popular Data Distribution Service (DDS) publish-subscribe middleware by a subscriber-centric software caching feature, which is then supported by an appropriate network scheduling. Using realistic data sets and ML applications from the popular Autoware benchmark and a zonal architecture according to the P802.1DG automotive Time-Sensitive Networking (TSN) network profile, we demonstrate that both cabling structure and network latency can be significantly reduced for both 1 Gbps and 10 Gbps TSN technologies. This result enables faster perception and/or lower cable cost, at no loss in data quality or reliability