The “Viterbo2025-1.0” Dataset for Training and Testing AI Algorithms for Criminal Actions Detection and Recognition

Abstract

The recent escalation of urban street violence is requiring the development of effective and timely detection strategies. Among the most promising approaches is the use of artificial intelligence-based human action recognition (HAR) algorithms. However, despite substantial advancements, current HAR systems exhibit limited reliability when dealing with criminal actions, especially in terms of high false alarm rate. This is primarily due to the unsuitability of the existing training datasets, which are generally derived from surveillance cameras (poor video quality) or from movie scenes (too artificial and too similar among each other). To overcome this issue, the “Viterbo2025-1.0” dataset, has been created. It is based on a novel approach, where criminal actions are performed by trained actors in authentic urban locations and recorded, by a group of off-the-shelf cameras, mounted on streetlamps, which could be used to set up a cheap yet effective city video surveillance system. The dataset includes three action classes-frontal assault, rear assault, and riot-recorded across nine distinct urban environments, and is thought to ensure broad variability and authenticity, enabling both model generalizability and adherence to reality. Each video has then been tagged both temporally and spatially, indicating the precise timing and location of the violent act within each frame. The dataset is thus suitable for both training and benchmarking, also facilitating comparisons across AI models that use standard and advanced evaluation metrics. Preliminary results show that a state-of-the-art architecture achieved a higher F1-score (against a comparable accuracy) when trained with the proposed dataset then with the existing ones. Designed to be publicly accessible and extensible, the Viterbo2025-1.0 dataset aims to become a reference standard for future research in violence detection, offering a robust foundation for developing generalizable and accurate AI systems in the field of urban crime surveillance.