Abnormal Human Action Recognition Using MobileNetV2 with Advanced Data Augmentation and Fine-Tuning Techniques

Abnormal Human Action Recognition (AbHAR) encounters numerous obstacles, such as variations in viewing angles, occlusions, fluctuations in execution rates, camera movements, and intricate backgrounds. The presence of ambiguous environments, dynamic backgrounds, and adverse weather conditions intensifies these challenges, thereby complicating the task at hand. Recent advancements in deep learning and computer vision have facilitated the autonomous learning of various features by models, thereby enhancing AbHAR systems. This study introduces an AbHAR system utilizing MobileNetV2 architecture, specifically optimized for mobile and embedded systems. Through the optimization of MobileNetV2, the system can identify actions including headache, stomach-ache, chest pain, and coughing from video frames. The model adjusts to environmental variables by considering alterations in illumination, perspective, and backdrop. The study attained a remarkable classification accuracy of 97% through the implementation of optimal strategies in dataset management, which encompassed frame extraction, resizing, normalization, and data augmentation. The effective adjustment of MobileNetV2, combined with thorough training and assessment methodologies, illustrated the system's capability in precisely and consistently identifying atypical human behaviors.