Human activity recognition is a dynamic area within artificial intelligence. It involves identifying human actions during everyday tasks such as standing, sitting, and walking. One application of this technology is in supermarkets, where it can analyze consumer behavior or function as a surveillance tool to prevent theft. This particular study utilizes OpenPose and Convolutional Neural Networks (CNN) with a custom-collected dataset. The program detects human skeleton shapes from camera footage and classifies these shapes using CNN with the ResNet50 model, subsequently displaying the identified activities. The classified activities include standing, walking, picking up items, looking at items, and pushing a trolley. The testing results indicate a training accuracy of 99.76% and a validation accuracy of 96.52%, along with an accuracy score of 96.52%, a precision of 96.59%, a recall of 96.525, and an F1-score of 96.53%.
Copyrights © 2024