Problem

Manual monitoring of CCTV footage for suspicious activity is time-consuming and requires constant human attention.

Approach

Collected and annotated a surveillance dataset, extracted temporal features from frame sequences with OpenCV, and trained an LSTM in TensorFlow/Keras to classify short clips as normal, shoplifting, robbery or explosion on an imbalanced dataset. Deployed via Streamlit with automated incident reporting.

Results

92% accuracy at 2.5 seconds per video segment, with an automated report for every flagged incident.