Lightweight Edge Intelligence Empowered Near-Crash Detection Towards Real-Time Vehicle Event Logging

Abstract

A major role of automated vehicles is that vehicles serve as mobile sensors for event detection and data collection, which support tactical automation in autonomous driving and post-analysis for traffic safety. However, most data collected during regular operations of vehicles are not of interest, while it costs a large amount of computation, communication, and storage resources on the cloud servers. Vehicular edge computing has emerged as a promising paradigm to balance these high costs in traditional cloud computing. But edge computers often have limited resources to support the high efficiency and intelligence of advanced vehicular functions. Motivated by the existing challenges and new concepts, this paper proposes and tests a lightweight edge intelligence framework for vehicle event detection and logging that runs in an event-based and real-time manner. Specifically, this paper takes vehicle-vehicle and vehicle-pedestrian near-crashes as the events of interest. The lightweight algorithm design of modeling the bounding boxes in object detection/tracking enables real-time edge intelligence onboard a vehicle; The event-based data logging mechanism eliminates redundant data onboard and integrates multi-source information for individual near-crash events. Comprehensive open-road tests on four transit vehicles have been conducted.