PosturalNet, an Advanced Architecture to Detect Shoplifting at Megastores

Human activity recognition (HAR) requires a full understanding of postural dynamics to create sophisticated surveillance systems that specifically detect shoplifting activities. PosturalNet represents a new architectural framework that uses human postural analytics to detect shoplifting incidents in large retail spaces. The proposed method uses a 2D body joint detection algorithm named OpenPose to extract postural information. Subsequently, these postures are pre-processed, and then features such as joint velocities, normalized body joints, magnitudes and angles are extracted. A combination of feature encoding techniques together with dimension reduction strategies maintains vital information while improving system computational efficiency. Testing PosturalNet on synthetic shoplifting data yielded an accurate detection rate of 96.87%. The method demonstrates robust performance in real-time applications which handle occluded environments when applied to practical store surveillance footage.