Semantic Editing of Traffic Near-Miss and Accident Dataset Using Tune-A-Video

Abstract

Developing effective traffic monitoring systems for accident detection relies heavily on high-quality, diverse datasets. Many existing approaches focus on detecting anomalies in traffic videos. Still, they often fail to account for how varying environmental conditions, such as time of day, weather, or lighting, might influence the occurrence of near-misses or accidents. In this study, we explore the potential of Tune-A-Video to apply semantic editing techniques to an existing traffic near-miss and accident dataset. By modifying the visual environment, such as changing the time of day, weather, or lighting, we aim to generate realistic footage variations without altering the core events like near-miss incidents or accidents. This method enhances the dataset with more varied and realistic traffic conditions, improving its representativeness of real-world scenarios. The primary objective is not to create a new dataset but to assess the impact of semantic editing on the dataset’s diversity and its effect on model performance. The results show that using Tune-A-Video for semantic editing can enrich the dataset, making it more suitable for training machine learning models. This approach helps improve the accuracy and robustness of computer vision models, particularly for traffic monitoring and accident detection applications, offering a promising tool for traffic safety systems.