Enhanced Rear-End Collision Detection and Localization Scheme Using LSTM, DRNN, and DNN in Fog-Based Internet of Vehicles

Previous studies were attributed to poor detection of rear-end collision due to high error rate, poor model design, inaccurate collision detection and poor model convergence rate. Other factors influencing collision in the IoV were not considered in the previous literatures, the previous works also rely solely on data-processing software to determine the level of influence that these factors have regrading collision occurrence, run simulation using PTV Vissim traffic simulator to generate the datasets needed for model training and testing to detect collision. The previous approaches were time consuming, incomplete and lack robustness. The objectives of these studies are to determine the relationship between collision features and collision occurrence directly using the models rather than relying on outcome of data-processing software, we also introduce two new features (number of interacting vehicles and junctions) that were not considered in the previous literatures and enhance the performance of the previous work. In our study, algorithms in the previous literature are enhanced, apply the algorithms to directly relate features with rear-end collision, introduced localization to determine the area where collision event occurs. Our proposed scheme is based on single and hybrid modes involving LSTM, DRNN-LSTM, and DNN-LSTM. The MSE and RMSE were reduced from 0.000010347 to 0.0000000019433 and 0.0032166 to 0.000044082, respectively, while the AES achieved was from 87 s (418us/step) to 70secons (38 ms/step), the accuracy of the proposed schemes were also determined where each model attained 100%.