SentinelPose: A Temporal Consistency-Aware Framework for Pose-Based Behavioral Anomaly Detection in Retail Surveillance

Retail stores widely use CCTV cameras to monitor customer activities and maintain security.However, continuously watching multiple video feeds is difficult for security staff, and unusual activities may sometimes go unnoticed.Existing automated surveillance methods can help identify suspicious movements, but their performance may be affected by inaccurate body pose estimation, tracking errors, and normal customer actions that appear unusual.To address these challenges, this paper proposes SentinelPose, a reliability-aware framework for detecting unusual human behavior in retail environments.The proposed system uses YOLO to detect people in CCTV footage and ByteTrack to track their movements across video frames.MediaPipe Pose is used to extract body landmarks, which are organized into time-based sequences and analyzed using an LSTM Autoencoder.The model learns common movement patterns and identifies possible anomalies based on reconstruction error.To improve the reliability of alerts, the proposed approach considers the quality of extracted pose information and checks whether unusual movements continue over multiple frames before generating an alert.When a potential anomaly is identified, the system is designed to save relevant video evidence and display the alert on a monitoring dashboard for security staff to review.The framework is intended to reduce unnecessary alerts and support efficient surveillance without using facial recognition or automatically identifying anyone as a shoplifter.The proposed method will be evaluated using detection precision, recall, false alarm rate, and processing speed.