A MULTIRESOLUTION CNN FRAMEWORK FOR REAL-TIME DETECTION AND CLASSIFICATION OF HUMAN ACTIVITIES

Abstract

Accurate and real-time recognition of human activities is fundamental to the advancement of intelligent surveillance, human–computer interaction, assisted living, and smart IoT environments [1], [5]. Traditional approaches rely on handcrafted features and shallow machine-learning classifiers, which struggle to capture the multi-scale spatial and temporal dynamics of complex human movements, leading to poor generalization in real-world scenarios [4]. To address these limitations, this research proposes a multiresolution Convolutional Neural Network (CNN) framework designed to detect and classify human activities from video sequences with high precision and low latency [12], [18]. The model incorporates multi-scale feature extraction layers that learn hierarchical motion representations from fine-grained postures to coarse full-body dynamics, enabling robust activity recognition even under variations in lighting, camera angles, background clutter, and occlusion [7], [6]. A parallel multi-resolution branch architecture processes both high- and low-resolution spatial maps to capture micro-level motion cues and macro-level behavior patterns, while an adaptive feature fusion module aggregates complementary features for effective decision-making [12], [18]. Extensive experiments on benchmark datasets— including UCF101, HMDB51, and Kinetics-Human—demonstrate that the proposed framework surpasses conventional CNN and recurrent architectures in terms of accuracy, detection speed, and robustness [5], [13]. The system achieves real-time inference on edge-GPU platforms, making it suitable for deployment in smart surveillance systems, autonomous robotics, and elder-care monitoring applications [15]. Overall, the multiresolution CNN framework represents a scalable and efficient solution for high-performance human activity recognition in dynamic and resource-constrained environments [7], [15].