Inferring Human Driver Intent in Partial Deployment of Connected Autonomous Vehicles: The Lane Change Case

Co-existence between autonomous vehicles (AVs) and human-driven vehicles expected in the next few decades poses a problem for AVs to infer human drivers' intents nearby and cope with them safely and efficiently. To address this issue, we develop a light-weight deep learning model for a connected autonomous vehicle (CAV) to infer intents in a safety-critical case of lane changes made by human-driven vehicles. Through experiments with the real trajectory dataset NGSIM, we show that a simple Multi-Layer Perceptron (MLP) model can predict lane change events with high accuracy comparable with more sophisticated models. The model is intentionally designed to work with the simplest 3-vehicle topology to foster real-time execution on the resource-constrained computing platforms on AVs. Still, the model achieves 85% accuracy over 5 to 8 seconds prediction horizons so that AVs can have enough time to prepare for an upcoming lane change event.