Zexuan Jia

Autonomous driving significantly benefits from data-driven deep neural networks that require large-scale and high-quality datasets for training. Several well-known datasets such as nuScenes, CADC, and SUSCape have been applied to train deep models for autonomous driving. However, these datasets generally have many redundant and low-quality frames that take extra training time and affect the performance of the trained models. In this work, we propose a novel three-layer method by considering scene complexity, instance quality, and sensing uncertainty to remove redundant data and maintain valuable data for optimizing autonomous driving datasets with smaller data sizes and equivalent effectiveness. We apply the method to optimize four well-known datasets nuScenes, CADC, SUSCape, and Carla-4Scenes. Furthermore, we retrain the well-known perception algorithms on these optimized datasets. The experimental results show that these retrained perception algorithms on the optimized datasets perform equivalent even better accuracy with less training time. We conclude that our three-layer method successfully optimizes autonomous driving datasets with smaller sizes and equivalent effectiveness, which contributes to faster training with lower hardware requirements.

Long-horizon traffic prediction is critical for automated vehicles control near intersections and queue-forming road segments, enabling energy-efficient and congestion-aware planning. This work proposes a control-oriented representation of future longitudinal motion constraints as a virtual leader acting on the ego vehicle, and develops a hybrid framework for virtual-leader prediction. In the short term, a deep learning model predicts future occupancy and flow, from which a most likely collision-free virtual leader constraint is inferred. In the long term, a queueing-theory-based model extends this constraint based on the traffic stage estimated from the short-term prediction. On the Waymo Motion Dataset, the short-term predictor achieves an area under the curve (AUC) of 0.755 for occupancy prediction and an endpoint error (EPE) of 5.862 for flow prediction, using 4s of historical data and a 5s prediction horizon. For full-horizon virtual leader prediction, the proposed framework reduces the average displacement error (ADE) and final displacement error (FDE) by 18.7% and 38.4%, respectively, compared with a queueing-theory-only baseline. These results support the proposed approach as a promising framework for control-oriented virtual-leader-based motion constraint representation in complex traffic environments.

Autonomous driving significantly benefits from data-driven deep neural networks that require large-scale and high-quality datasets for training. To achieve higher performance or conduct critical testing, trajectories with game-theoretic properties are necessary. However, defining such scenarios and collecting large-scale data from reality or simulation is hard and costly. In this work, we propose a novel two-step method that generates traffic trajectories with game-theoretic properties, based on the Large Language Model (LLM). We apply the method to generate a challenging traffic trajectory dataset, which is validated on many downstream tasks, including perception and prediction. The experimental results show that the generated trajectories can achieve high-difficulty testing and advanced training. We conclude that our method successfully generates the challenging traffic trajectories, which contribute to collecting such critical data.