Enhancing Highway Safety with LSTM-based Vehicle Intention Prediction (HVIP)

Predicting vehicles' motion on highways has become crucial for enhancing road safety and traffic flow. Deep learning, which reached exceptional results in various applications, is now the leading approach for vehicle motion prediction. This work presents a deep learning-based model using Long Short-Term Memory (LSTM) networks to classify vehicle intentions into lane-keeping, left lane-changing, and right lane-changing. Utilizing the PREVENTION dataset, which provides naturalistic driving data, a sequence of centre points, longitudinal distances, lateral distances and yaw angles of a vehicle has been extracted to train the model to predict lane changes effectively. The model consists of three LSTM layers, a dense output layer and a drop out added between layers to prevent overfitting. Experiments focusing on optimizing the model parameters, learning rate, sequence length, and batch size, were conducted to determine their impact on prediction effectiveness. The best-performing model showed significant results in lane changing prediction, achieving an accuracy of 84%, precision of 89%, and recall of 82%. Future work will aim to train the model on more complex features, to account for vehicle inter-dependencies and interactions, and test model on various highway scenarios for a more reliable system.