Pose-Based Suspicious Activity Detection Using YOLOv12n and MediaPipe in Smart Surveillance Systems

Abstract

Detecting suspicious activity in video surveillance is critical for public safety. Early identification of pre-incident behaviors, such as "looking around" nervously, can prevent crimes before they occur. However, traditional surveillance relies on continuous human monitoring, which is prone to error and resource-intensive. Existing automated methods struggle with real-time stability and produce flickering predictions. We present a two-stage pose-based detection system: YOLOv12n detects humans in each frame, and MediaPipe Holistic extracts pose features. A Support Vector Classifier (SVC) and Neural Network (NN) then classify 15-frame temporal segments as either "still" or "looking around." On our custom dataset, SVC achieves 97.35% accuracy while NN achieves 97.00%. Temporal smoothing eliminates prediction flickering, enabling stable real-time deployment. SVC provides better precision-recall balance, NN offers greater adaptability to complex pose variations.