This paper explores imitation learning, which mimics the behavior of expert drivers for autonomous vehicles (AVs) to learn to navigate complex traffic environments. Existing imitation learning frameworks often focus on expert demonstrations, overlooking the potential for utilizing complex driving data from surrounding traffic participants. This study presents a data augmentation strategy that utilizes observed trajectories of nearby vehicles captured by AV sensors as additional demonstrations. A simple vehicle-selective sampling and filtering strategy prioritizes informative and diverse driving behaviors, resulting in a richer training dataset. Evaluation results using a representative learning-based planner on a large real-world dataset demonstrate improved performance in complex driving scenarios. Specifically, the system reduces crash rates and improves safety metrics, achieving performance comparable to or exceeding that achieved using the full dataset even with only 10% of the original dataset. Through ablation studies, we analyze the selection criteria and demonstrate that simple random selection can result in poor performance. This study highlights the value of utilizing diverse real-world trajectory data in imitation learning and provides insights into data augmentation strategies for autonomous driving.
