An Enhanced Depth Motion Map Framework for Accurate Human Action Recognition Using Deep Learning

Human Action Recognition (HAR) using depth sequences has become increasingly important due to its robustness against illumination changes and background variations. However, conventional Depth Motion Maps (DMMs) often accumulate noisy and redundant motion information, reducing recognition accuracy. This paper proposes a Deep Learning–Driven Enhanced Depth Motion Map (EDMM) framework for accurate human action recognition. The proposed approach introduces a pixel-level motion analysis mechanism combined with a spatial window-based motion significance strategy to effectively suppress false motion regions and depth noise. The refined motion maps are then processed using a customized Convolution Neural Network (CNN) to learn discriminative spatial–temporal features for classification. The framework is evaluated on benchmark datasets such as MSR Action 3D and UTD-MHAD. Experimental results demonstrate improved recognition performance compared to traditional DMM-based and handcrafted feature methods. The proposed EDMM-CNN model enhances motion representation quality while maintaining computational efficiency, making it suitable for real-time applications in surveillance, healthcare monitoring, and human–computer interaction.