Real-time Tra ic Estimation at Vehicular Edge Nodes

Tra c estimation has been a long-studied problem, but prior work has mostly provided coarse estimates over large areas. This work proposes e ective ne-grained tra c volume estimation using invehicle dashboard mounted cameras. Existing work on tra c estimation relies on static tra c cameras that are usually deployed at crowded intersections and at some tra c lights. For streets with no tra c cameras, some well-known navigation apps (e.g., Google Maps, Waze) are often used to get the tra c information but these applications depend on limited number of GPS traces to estimate speed, and therefore may not show the average speed experienced by every vehicle. Moreover, they do not give any information about the number of vehicles traveling on the road. In this work, we focus on harvesting vehicles as edge compute nodes, focusing on sensing and interpretation of tra c from live video streams. With this goal, we consider a system that uses the dash-cam video collected on a drive, and executes object detection and identi cation techniques on this data to detect and count vehicles. We use image processing techniques to estimate the lane of traveling and speed of vehicles in real-time. To evaluate this system, we recorded several trips on a major highway and a university road. The results show that vehicle count accuracy depends on tra c conditions heavily but even during the peak hours, we achieve more than 90% counting accuracy for the vehicles traveling in the left most lane. For the detected vehicles, results show that our speed estimation gives less than 10% error across diverse roads and tra c conditions, and over 91% lane estimation accuracy for vehicles traveling in the left most lane (i.e., the passing lane).