Abstract:
Monitoring large amounts of video or image data for violence can be challenging and prone to human error. This project aims to solve this problem by creating an automated system designed to detect violent anomalies in both images and videos. The system reduces the need for constant human monitoring and ensures faster, more accurate identification of violent activities. It can be applied across various areas, including surveillance, public safety, and other fields where detecting violence is critical. We trained two models for image classification: a Custom CNN with 89% accuracy and ResNet50 with 95% accuracy. For generating descriptions of detected violent events, we use ViT (Vision Transformer) for image analysis and GPT-2 for caption generation. These models work together to automatically detect and describe violent actions in real time. By leveraging deep learning techniques, our system ensures high accuracy in identifying violent activities. The goal is to make violence detection more efficient, reducing human error and enabling quicker responses to potential threats.
