Real-Time Detection of Criminal Actions in the Everyday Life, From Camera-Equipped Streetlamps

Timely identification of illicit activities stands as a primary issue to be addressed, demanding swift and effective responses from emergency services. Technological advances in computer vision and image processing allowed industry and research practitioners towards the implementation of automated digital technologies to perform such an important task. In light of it, we introduce a system designed to detect and recognize criminal activities in a urban environment, through video analysis, employing advanced deep learning methodologies. The videos will be obtained from recordings cameras installed on the streetlamps of a city, allowing a wide and capillary monitoring. This paper is focused specifically on the algorithm we designed to identify criminal actions, which is based on the You Only Look Once (YOLO) algorithm. The algorithm has been trained on a public dataset, commonly used for anomaly detection in city contexts. The dataset has been prepared and labeled specifically to suit our purpose, considering two different labeling strategies. The evaluation of the proposed model yielded promising results, demonstrating the effectiveness of the approach and laying the groundwork for future enhancements.