Deep learning based trajectory prediction system for autonomous vehicle using heterogeneous datasets overcoming data defection

Abstract

Despite globally overwhelming research and development efforts toward autonomous vehicle technologies, there are a lot of difficulties in the real-world driving task. Even though several autonomous vehicles are in op-eration in public, human intervention to overcome difficult driving scenes is frequently required. Such a situation is originated from the difficulties in future uncertainty due to the variability of the driving characteristics of the human-driver and inter-vehicular interaction among multiple agents. Therefore, a probabilistic precise future prediction algorithm is essential for reliable and robust autonomous driving. In this dissertation, a methodology for overcoming the various data defections in predicting the future trajectory of the vehicles is proposed. There-fore, the general technical difficulties in predicting the future of vehicles originating from the data quality are overcome so that the broader and detailed context of the vehicle's future maneuver can be predicted with higher performance. The proposed prediction framework extracts the maneuver context of the vehicles from the infinitely diverse driving scene and predicts the reaction change of the surrounding vehicles for various future plans of the ego vehicle. Here, transfer learning based on the heterogeneous dataset and a semi-supervised learning approach is employed to overcome the data imbalance and data deficit problem. The proposed framework consists of: 1) a Scalable trajectory prediction system using Graph Neural Network (GNN) capable of dynamically varying traf-fic scenes, 2) a Representation learning framework for extracting the maneuver context of the vehicles securing the generality in terms of the vehicle maneuver, and 3) Reactive trajectory prediction for capturing reaction changes of the surrounding vehicles for a maneuver plan of the ego vehicle. The proposed trajectory prediction network is a total trajectory prediction solution capable of the diversity of the driving environment and vehicle maneuver. Thereby, the travel efficiency, as well as the safety, can be secured by integrating with the path plan-ner for the autonomous vehicle.