Crime Prediction by Detecting Violent Objects and Activity Using Pre-Trained YOLOv8n and MoViNet A0 Models

Crime prediction is crucial to contemporary law enforcement and policies ensuring public safety. This study introduces a novel method for crime prediction by using deep learning. Specifically, the methodology involves the detection of violent objects and activities via the use of pre-trained YOLOv8n and MoViNet a0 models. To achieve the goal, two deep learning models have been considered. YOLOv8n, the initial model, has been fine-tuned for violent object detection using a weapon detection dataset. The second model, MoViNet a0 base, has been fine-tuned for violence activity detection on a hockey battle dataset. This paper explores the fusion of these models to enhance crime prediction by identifying violent objects and activities in real-world scenarios. With an accuracy of 91.7% for recognizing violent objects and 94% for detecting violent activity, the suggested approach provides an efficient solution for crime prediction by detecting violent objects and activity. Experimental findings reveal that the models can reliably recognize violent items and behaviours. Through object detection and action recognition, this technique provides law enforcement and security professionals with a comprehensive view of possible risks, enabling proactive crime prevention.