Real-time abnormal activity detection in surveil-lance systems requires a balance between accuracy, computa-tional efficiency, and privacy preservation. This paper presents a lightweight and efficient framework for detecting human activities using skeletal pose estimation on standard consumer-grade hardware. The proposed system extracts 33 human body landmarks and converts them into normalized spatial feature vectors using a centroid-based approach to achieve translation and scale invariance.

For classification, a K-Nearest Neighbors (KNN) model is employed with a visibility-weighted Manhattan (L1) distance metric to improve robustness against noisy or partially occluded landmarks. Additionally, an adaptive distance threshold derived from intra-class variations and a temporal smoothing mechanism are introduced to enhance prediction stability and reduce false detections.

The system supports multi-person detection using region-based processing and integrates an automated alert mechanism that generates real-time notifications, including snapshot capture and email transmission via SMTP. Experimental results demonstrate that the proposed framework achieves high accuracy with low latency, making it suitable for real-time surveillance and safety-critical applications.

Index Terms—Pose Estimation, K-Nearest Neighbors (KNN), Anomaly Detection, Real-time Surveillance, Multi-person Detec-tion, Edge Computing