iTA: Inferring Traffic Accident Hotspots with Vehicle Trajectories and Road Environment Data

Traffic accidents, like rear-end collisions and vehicle-to-pedestrian accidents, bringing serious threats to traffic management and people’s lives and property. Existing methods rely on historical traffic accident data and are not suitable for expanding cities. Against this background, we propose a framework to infer urban traffic accident hotspots by vehicle trajectory data and heterogeneous data. Firstly, we extract parking events from vehicle trajectories leveraging a map matching algorithm and obtain the parking hotspots under different weather and time periods by semantic segmentation and clustering. Secondly, we extract three sets of contextual features of parking hotspots, i.e., road names, POIs, and traffic flow. Finally, leveraging a graphical convolutional network, we build a model to infer whether a parking hotspot is a traffic accident hotspot based on contextual features. We evaluate our model leveraging real datasets in Xiamen. Results prove that our model accurately infers traffic accident hotspots in Xiamen with F1-Score of 0.794.