GuardianVision

"Guardian Vision" is a deep learning-based intelligent surveillance system designed for real-time anomaly detection in CCTV footage, encompassing video, audio, and image analysis. The system aims to detect abnormal events such as violence, presence of weapons, and accidents through advanced media processing and a robust alert mechanism. For video-based violence detection, the system utilizes InceptionNet for spatial feature extraction and GRU (Gated Recurrent Units) for temporal sequence classification, achieving an expected accuracy of over 90%. In the event of detecting violence or weapon presence, the system captures a screenshot from the video and sends an email alert with the evidence attached. For audio anomaly detection, the system uses the publicly available Audio-Based Violence Detection Dataset from Kaggle. Based on feasibility, either CNN or LSTM architectures are applied to classify violent audio cues. While audio-based detection may have variable accuracy due to noise and environment conditions, it serves as a supplementary channel to enhance system reliability. Additionally, the system integrates image-based accident detection, using MobileNet for lightweight and efficient feature extraction from images, classifying scenes into “Accident” or “No Accident” categories. This multi-modal detection pipeline—video, audio, and image—ensures a comprehensive monitoring solution. Keywords: Anomaly Detection, MobileNet, InceptionNet, GRU, LSTM, CNN, Video Surveillance, Audio Violence Detection, Accident Detection, Email Alert System, Deep Learning, Media Analysis, Guardian Vision.