DMDAT: Diffusion Model-Based Data Augmentation Technique for Vision-Based Accident Detection in Vehicular Networks

Generative AI (GAI) has garnered significant attention recently with the upsurge of ChatGPT and similar models. Research has shown that GAI technology has significantly contributed to several domains like healthcare, cybersecurity, industry 4.0, gaming, etc. GAI technology is also vital in vehicular networks and intelligent transportation systems. This paper presents a diffusion model-based data augmentation for enhanced vehicular network applications. The existing data augmentation techniques fail to modify the semantics of the images in the dataset. Addressing this, we propose a diffusion-based model for effective data augmentation. The augmentation technique involves feeding a randomly generated prompt (covering several image features like weather, lighting, background, visibility, and surface) along with canny edge features of an existing image to a diffusion model to generate an image with desired features. We used a vision-based accident detection dataset and a MobileNet-based deep learning model to understand the efficacy of the proposed data augmentation technique. Our experimental analysis indicates the proposed data augmentation technique improves the performance notably (3%).