Intelligent Driving Risk Analysis using Real-Time Transformer-GRU-BiLSTM Network for Vehicle Detection

Nowadays, intelligent transportation systems continue to improve and extend risk analysis in real-time driving scenarios, which has become more critical for increasing road traffic safety. This study proposes an Intelligent Driving Risk Analysis (IDRA) framework for vehicle detection and risk assessment using a hybrid Transformer-GRU-BiLSTM (TGBNet) model. Specifically, the proposed model combined the Transformer architecture for modeling long-range dependencies between successive driving data and the GRU and BiLSTM layers for temporal feature extraction and contextual understanding, respectively. The proposed approach combines deep learning (DL) algorithms such as CNN, RNN, Autoencoder, etc., to detect vehicles in active environments and evaluate risk connected with the ability to drive depending on the speed of cars, static/moving objects, and distance of object with the vehicle like sensor fusion from LiDAR, camera, and radar data, leading to improved detection accuracy. The two datasets, KITTI and Waymo, were collected from online sources like Kaggle. Experimental results show that TGBNet significantly outperforms traditional CNN-LSTM and single RNN-based models. Finally, the proposed TGBNet evaluates the performance in terms of MSE of 0.022, EUE of 0.22, and CPE of 23.2% for KITTI. For Waymo’s dataset, the performance is MSE of 0.019, EUE of 0.21, and CPE of 21.4%, representing the driving scenarios into safe, moderate, and high-risk levels.