Real-Time Anomaly Detection in CCTV Surveillance

A deep learning system that watches surveillance footage and automatically flags threatening activity — theft, violence, and property damage — in real time.

---

What It Does

RealtimeSurviellance.mp4

Most CCTV systems record everything but catch nothing until a human reviews the footage after the fact. This project flips that — a model watches the stream live, classifies what's happening every few frames, and overlays the result directly on the video.

Four output classes:

• normal — no incident
• theft — shoplifting, robbery, burglary
• violence — assault, fighting, abuse
• property_damage — arson, vandalism

---

Architecture

Video Frame
    │
    ▼
CLAHE contrast enhancement          ← improves visibility in dark CCTV footage
    │
    ▼
YOLOv8 person detection + crop      ← focuses the model on people, not background
    │
    ▼
Top-16 motion frame selection       ← picks the most action-rich frames from a 64-frame buffer
    │
    ▼
ResNet50 feature extraction         ← pretrained CNN, outputs 2048-dim vector per frame
    │
    ▼
BiLSTM classifier                   ← learns temporal patterns across the 16-frame sequence
    │
    ▼
4-class prediction + smoothing      ← rolling window reduces single-frame flicker

The same preprocessing pipeline runs identically during training and live inference, so the model never sees a different distribution at runtime.

---

Results

Evaluated on a held-out stratified test split (15% of dataset).

Class	Precision	Recall	F1
normal	0.9518	0.9130	0.9320
theft	0.8955	0.9104	0.9029
violence	0.8530	0.9028	0.8772
property_damage	0.8889	0.8889	0.8889

Test Accuracy : 90.72%

Applications

• Security operations — automated pre-screening across multi-camera feeds
• Retail loss prevention — real-time theft flagging
• Smart city CCTV — large-scale public incident detection
• Forensic review — fast search through archived footage

---

Tech Stack

Role	Tool
Person detection	YOLOv8n (Ultralytics)
Feature extraction	ResNet50 (Keras, ImageNet)
Temporal model	Bidirectional LSTM (TensorFlow)
Video decoding	Decord
Frame processing	OpenCV
Motion scoring	PyTorch

---

Dataset

UCF Crime Dataset — real-world CCTV footage across 13 anomaly categories, consolidated into 4 classes for this project. Class imbalance is handled through per-class window oversampling and weighted loss during training.

---

Kaggle notebook project — GPU (P100/T4) recommended for feature extraction and training.