Infrastructure-based Detection and Localization of Road Users for Cooperative Autonomous Driving

Abstract

Driving is an activity that can become extremely and unpredictably dangerous within a few seconds, and most often difficulties arise in complicated scenarios, such as intersections, roundabouts, or high traffic areas. Detection and localization of all road-users (vehicles, pedestrians, etc.) present at any given time could greatly reduce risk and accident rates by predicting and warning participants of potentially dangerous situations before they occur. This is a difficult task to do well from a single on-road perspective (a vehicle’s on-board sensors), but roadside units mounted in the infrastructure can provide a few distinct advantages. We can obtain a top-down, unobscured view of the intersection with a static background. A static background leads to a simpler learning task and allows models to be trained on smaller datasets. Additionally, results (and warnings) could be communicated to all road-users, essentially allowing many users to “share” very powerful computational resources in difficult environments, rather than requiring each vehicle to have its own resources that are likely excessive in most simple driving scenarios. This project describes detection results from roadside cameras and presents a 3D localization method for achieving sub-half-meter accuracy for vehicles and pedestrians moving through the intersection. We use a convolutional neural network (CNN) for detecting road users within an image and use an indexed location map to obtain localization information.