Predicting Road Traffic Risks with CNN-and-LSTM Learning Over Spatio-Temporal and Multi-Feature Traffic Data

Offering traffic safety information to drivers and passengers is one of essential services towards the smart city. Recent research utilizes AI models to analyze the collection of IoT-driven data in transportation environments. Exploring unveiled characteristics of traffic information to improve traffic control and accident prevention on roads, this way becomes plausible. Prior studies exploited various sorts of spatio-temporal traffic data to achieve the traffic prediction using deep learning models. Without understanding the complexity of spatio-temporal data, however, their efforts have not fully shown the effectiveness of deep learning-based traffic prediction and risk presentation. In this paper, our study first applies the Pearson correlation coefficient to clarify that traffic accidents appear in high correlation with time and space patterns. We identify multiple features from traffic domains, and employ CNN first and then LSTM learning techniques on several volumes of spatio-temporal traffic data, including weather, time, traffic flow, and historical traffic accidents and locations, etc. Our study shows that the combination of CNN and LSTM learning on spatio-temporal traffic data is applicable and useful for traffic risk prediction. Under experiments and demonstrations with actual traffic datasets, our proposed traffic risk prediction scheme, called CLwST, can exhibit more accurate results, faster convergence and lower loss in comparison with the two prior studies based on LSTM and ConvLSTM schemes.