Human Action Recognition Using Depth Motion Images and Deep Learning

Human activity recognition (HAR) aims to recognize actions performed by the subjects and the environmental conditions. It is essential for various computer vision applications that require insights into human behavior, such as virtual surveillance, human–computer interface, and robotics. In this paper, we proposed a method for human action recognition from Depth Motion Images (DMI), which maps sequences of depth maps of RGB videos generated from the MSRAction3D public dataset. Our stated deep neural network architecture, based on convolutional neural networks (CNNs), can identify 20 different types of actions. We also evaluated the model's performance using various evaluation metrics, including Precision, Recall, F1-Score, ROC curve, and AUC. The model demonstrated its ability to make precise predictions with a Precision of 84.19% and effectively recall actions with an 82.47% Recall, resulting in an 81.96% F1-Score during the 50th training epoch with the highest AUC value of 0.93. Our method obtains an accuracy of 82.81% compared to the state-of-the-art approaches currently in use and is based on Depth Map techniques.