YOLOv5 for Road Events Based Video Summarization

In recent years, metropolitan cities have seen a rapid rise in vehicular traffic, thereby increasing the risk of accidents. Automatic detection of accidents will shorten the response time for timely treatment and justice for the victims. Accordingly, this paper utilizes YOLOv5 for road events (traffic accident) detection and also proposes an event-based video summarization technique for reducing massive surveillance data storage. With the enforcement of privacy laws, it is challenging to get publicly available real-world road surveillance videos for training computer vision models. To cope with the scarcity of real-world data, this paper utilizes synthetic road surveillance data to train a YOLOv5 model for detecting road events from surveillance videos. The synthetically trained model is then tested on real-world road traffic surveillance videos. For experiments, the synthetic dataset is divided into 60% training, 20% validation, and 20% test sets. The proposed approach detects accidents across video frames and generates video summaries centered around the accident for faster future visual data analytics. Experimental results prove the efficacy of the developed scheme to achieve a reduction of storage space and the duration of summarized videos by a range of 20–50% for test videos.