SafeVision: AI-Powered Surveillance System

Abstract

Conventional closed-circuit television (CCTV) surveillance infrastructure operates almost entirely in a reactive mode: cameras record footage and human operators decide, after the fact, whether anything significant occurred. This passive approach carries well-documented shortcomings. Research has established that human attention during continuous screen monitoring drops sharply after twenty minutes [14], making real-time threat detection from human observation alone an unreliable process at scale. SafeVision addresses this gap with an AI-powered, two-stage video anomaly detection system built on modern deep learning. Stage 1 screens every video clip for normal or anomalous activity. When an anomaly is detected, Stage 2 immediately classifies the type of incident into one of three super-categories: Fighting/Violence, Armed Threat, or Property Crime. Results are delivered through a polished, operator-facing web dashboard built with Next.js [18] and FastAPI [17]. The core of the system is a feature-extraction and temporal-reasoning pipeline. CLIP ViT-B/16 [4], a Vision Transformer pre-trained on 400 million image-text pairs, encodes each video frame into a 512-dimensional semantic embedding. A lightweight TemporalTransformer, a two-layer encoder-only transformer with sinusoidal positional encoding [6] processes sequences of sixteen frame embeddings to produce a single clip-level classification. Training used a unified dataset combining the UCF-Crime benchmark [1] (1,900 surveillance videos across 13 crime categories) and the Smart-City CCTV Violence Detection Dataset (SCVD) [2]. Stage 1 achieved a validation macro F1 score of 0.8950, supported by a macro precision of 0.90 and a recall of 0.90. Stage 2 achieved a macro F1 score of 0.7565, with a macro precision of 0.76 and a recall of 0.75 on three-class crime classification. The entire inference stack requires under 750 MB of VRAM, making it deployable on commodity hardware. Limitations include the absence of real-time RTSP streaming and frame-level temporal localisation, both identified as the highest-priority directions for future work.