Computer Vision for Live Map Updates

Abstract

Real-time traffic monitoring has had widespread success via crowd-sourced GPS data. While drivers benefit from this low-level, low-latency map information, any high-level traffic data such as road closures and accidents currently have very high latency since such systems rely solely on human reporting. Increasing the detail and decreasing the latency of this information has significant value. In this work we explore the idea of real-time crowd-sourced map updates from visual data. We propose a system that uses object detection to detect hazards which are then reported via 4G LTE to a local server on the edge. This edge server aggregates the data and relays updates to other vehicles inside its zone. We call our system LiveMap. We demonstrate detection accuracy on hazards and characterize the system latency. We propose and develop two extensions that can improve system functionality. The first improvement is a semantic change detection pipeline, which can detect changes between image pairs to provide high-level map updates as well as enable accurate removal of stale hazards. Finally we develop a novel visual odometry algorithm to improve hazard localization.