DeepRisk: A Deep Transfer Learning Approach to Migratable Traffic Risk Estimation in Intelligent Transportation Using Social Sensing

Abstract

This paper focuses on the migratable traffic risk estimation problem in intelligent transportation systems using the social (human-centric) sensing. The goal is to accurately estimate the traffic risk of a target area where the ground truth traffic accident reports are not available by leveraging an estimation model from a source area where such data is available. Two important challenges exist. The first challenge lies in the discrepancy between source and target areas (e.g., layouts, road conditions, and local regulations) and such discrepancy would prevent a direct application of a model from the source area to the target area. The second challenge lies in the difficulty of identifying all potential features in the migratable traffic risk estimation problem and decide the importance of identified features due to the lack of ground truth labels in the target area. To address these challenges, we develop DeepRisk, a social sensing based migratable traffic risk estimation scheme using deep transfer learning techniques. The evaluation results on a real world dataset in New York City show the DeepRisk significantly outperforms the state-of-the-art baselines in accurately estimating the traffic risk of locations in a city.