| Abstract: |
Maneuver prediction, especially lane change maneuver, is of critical importance for the safe navigation of autonomous vehicles. Although benchmark datasets exist for trajectory prediction, datasets specifically tailored for maneuver prediction are rare. This is particularly true for lane change prediction. To address this gap, in the present paper, an instrumented test vehicle is used to collect, process and label lane change maneuvers across various traffic scenes. The resulting dataset, referred to as WylonSet, consists of front-facing camera images, area-view camera images, vehicle state data and lane information. Thereby, over 400 driving sessions are collected and labeled, including approximately 500 lane change maneuvers, laying the foundation for our study. The main motivation behind this work is to analyze and predict lane change maneuvers for the ego-vehicle in urban traffic scenarios using deep learning models. In this study, a novel multi-modal deep learning architecture is proposed, comprising different modules to extract important features from the collected data. The visual module is built using Convolutional Neural Networks (CNNs) to capture features from all camera images, while the interaction module utilizes Graph Neural Networks (GNNs) to capture spatial features between detected entities in the traffic scene. The state module utilizes vehicle state data, while the lane module utilizes lane features. All these features are tracked in time using the temporal module of Recurrent Neural Networks (RNNs). The proposed architecture is trained and validated on WylonSet. Finally, the proposed learning architecture is implemented, and the resulting model for lane change prediction of the ego-vehicle is evaluated in different driving scenes and traffic densities. |