Abstract: Smart parking and traffic management are some of the major problems faced during the development of a smart city. There exist numerous challenges when trying to develop systems to solve the above-mentioned problem since Vehicular Ad-Hoc Networks (VANETs) produce a large volume of spatiotemporal data (vehicle speed, direction, and GPS position) in real-time. Nevertheless, due to their complexity, it is difficult to detect the existence of any abnormal events rapidly and reliably. In order to tackle such challenges, we suggest developing a hybrid deep neural network model, integrating Long-Short Term Memory (LSTM) network capable of capturing long-term dependencies and Squeeze-and-Excitation block (attention mechanism) highlighting important features. The proposed model is tested on AV-GPS-Dataset, containing GPS tracks of real vehicles as well as attack instances. Experiments show excellent results (accuracy equals to 99.81%, precision to 99.81%, recall to 99.73%, and F1 score to 99.74%), which outperform traditional methods significantly. Our results confirm that spatiotemporal data plays a vital role in solving problems stated.

Keywords: Vehicular Ad Hoc Networks (VANETs), Spatiotemporal Data, Anomaly Detection, Deep Learning, LSTM, Squeeze-and-Excitation, Smart Parking

Pages: 393-405