Trajectory prediction in lane-change vehicles with deep learning method

Abstract

In the rapidly advancing field of autonomous driving and advanced driver-assistance systems, accurately predicting vehicle trajectories during lane changes is a critical challenge. This research focuses on addressing this challenge by exploring different architectures for deep-learning based models capable of predicting the future paths of lane-changing vehicles, via features like vehicle’s velocity and their X and Y coordinates. Our models incorporate a combination of Temporal Convolutional Networks (TCNs), Bi-directional Long Short-Term Memory (BiLSTM) networks with attention mechanisms to capture the complex spatiotemporal dynamics of vehicle movements in diverse traffic scenarios. Other networks like Gated Recurrent Unit (GRU) and transformers are being studied as well. The models leverage real-world traffic data from the NGSIM dataset, integrating various features such as velocity, acceleration and lane positions as inputs to the model training. With a focus on both highway and urban environments, this approach aims to enhance safety and efficiency in autonomous driving systems by enabling more accurate, real-time decision-making in dynamic traffic conditions. Initial results demonstrate promising improvements in trajectory prediction accuracy, positioning the models as a possible advancement in the field of autonomous vehicle navigation.