Video Anomaly Detection (VAD) is an AI-powered violence classification system that analyzes video clips in real-time. Built on a MobileNetV2 + BiLSTM + Temporal Attention architecture, it detects and classifies four types of violent actions — Kick, Punch, Slap, and Group Violence — with up to 98% validation accuracy. The system extracts spatial features per frame using MobileNetV2, computes temporal difference features to capture motion dynamics, then passes the combined sequence through two Bidirectional LSTM layers and a Temporal Attention mechanism before classifying. Key highlights: Multi-clip inference (5 stride-based clips averaged) for robust prediction Focal loss training to prevent class collapse on minority classes Two-phase training strategy: feature learning then fine-tuning Streamlit web UI - upload any MP4/AVI/MOV file and get instant results