Advanced Human Activity Recognition Using Pose Estimation and Deep Learning Techniques

Human activity recognition (HAR) is a prominent research topic in computer vision. This topic has extensive applications in developing applications for human-machine interactions, monitoring, and various other fields. With the notable progress in research, several methods have been proposed to distinguish distinct types of human movements by utilizing color, depth, inertia, and skeletal data. To enhance the accuracy of human activity recognition and overcome the constraint of existing deep learning approaches that mainly rely on video frame inputs, we provide The solution being suggested using MediaPipe, a platform that provides pre-trained machine-learning models specifically designed for human pose estimation. These models will precisely identify and monitor essential body joints and movements while doing activities. Subsequently, the system will examine the identified poses to compute significant angles and distances for feature extraction. Following that, deep learning techniques such as Artificial Neural Networks (ANN) and 1D-convolutional Neural Networks (1D-CNN) are employed to categorize human behaviour. The proposed system's performance was assessed using our dataset to detect and categorize different human activities. The ANN and 1D-CNN models showed superior performance in recognizing human activity, with the ANN achieving an accuracy of 99.41 % and the 1D-CNN achieving an accuracy of 99.58 %.