Accurate Classification and Prediction of Remote Vehicle Position Classes Using V2V Communication and Deep Learning

Accurate classification and prediction of remote vehicle positions contribute significantly to the functionality of Advanced Driver Assistance Systems (ADAS). This essential information offers vital details about surrounding vehicles that are critically important in vehicle control decisions in order to avoid collisions and optimize vehicle performance. This paper investigates two deep learning neural network (DLNN) models: Long Short-Term Memory (LSTM) and Transformer-based deep neural networks, and effective features extracted from V2V communication signals for accurate detection and prediction of remote vehicle positions. Instead of predicting vehicle trajectories, this research focuses on detection and prediction of the remote vehicle positions characterized in 8 classes of locations immediately surrounding a host vehicle. These position classification and prediction results can be readily used by the host vehicle in making decisions involving lane change, making safe turns and overtaking, executing proper yielding, etc. We show through extensive experiments that the proposed DLNN models, LSTM and Transformers, are capable of effectively modeling the underlying dynamics of the 8 vehicle positioning classes, and providing situational awareness with the predicted remote vehicle positions. The experiments were conducted on V2V communication data collected from 69 real-world driving trips. Experimental results demonstrate that the proposed LSTM and the transformer-based DLNN systems outperform multilayer perceptron (MLP) systems by a large margin for both detection and prediction of remote vehicle position classes. The average prediction accuracy of the transformed-based DLNN systems using proposed geometric features combined with host and remote vehicle's location and speed outperformed the LSTM systems by more than 7%.