Link Travel Time Prediction from Large Scale Endpoint Data

Abstract

Existing systems for travel time estimation either use data collected from loop detectors and probe vehicle locations, or from GPS traces from cellphones of online users. The former methods of data acquisition are expensive, while the latter turns out to be infeasible in connectivity-poor regions. However, many crowdsourced taxi trip datasets (from Boston, Beijing, Rome, etc.) are publicly available which, despite containing limited information, can be made useful for inferring meaningful insights by certain amount of data engineering. The datasets are both cheap to acquire (hence available in large volumes), and impose less heavy connectivity requirements on the end user. One such crowdsourced dataset is the NYC (New York City) Taxi dataset, which contains only the end-point information for each trip. In this paper, a link (road segment) travel time estimation algorithm named Least Square Estimation with Constraint (LSEC) has been developed from such end-point data, which estimates travel time 20% more accurately than existing algorithms. The key idea is to augment a subset of trips with unique paths using logged distance information, as opposed to fitting adhoc route-choice models.