Vision transformer for detecting traffic congestion using image mapping

Leveraging advanced computer vision techniques, we present a novel method for analyzing and categorizing traffic videos by utilizing an image mapping technique combined with a UniVit (Unit Vision Transformer) model. First, video data is transformed into an image dataset, where each frame undergoes a mapping process based on vehicle detection using the YOLOv8s algorithm. These mapped images are then input into a vision transformer, which utilizes attention mechanisms to identify and analyze complex patterns across frames. The approach is evaluated on the UCSD (University of California San Diego) dataset, which features diverse weather conditions such as clear, overcast, and rainy scenarios, enabling a rigorous assessment under varied environments. Our method achieves a notable accuracy score of 99.89%, surpassing previous approaches on the UCSD benchmark. This improvement underscores the effectiveness of combining image mapping with advanced vision transformers, suggesting promising applications for intelligent traffic management systems aimed at mitigating traffic issues.