Human Body Posture Classification from Digital Images Using the InceptionV3 Architecture

Abstract

This study aims to develop a human body posture classification model based on digital images using the InceptionV3 architecture. The dataset used in this study is the MPII Human Pose Dataset, which contains a wide variety of human activities and body postures. The research began with label extraction from a metadata file in .mat format, followed by the selection of the 20 activity classes with the largest number of samples. Subsequently, the dataset was balanced using an undersampling technique, resulting in 140 images for each class. The images were then resized to 224×224 pixels, normalized using the preprocess_input function, and enhanced through data augmentation applied to the training set. The model was developed using a transfer learning approach with InceptionV3 as the base model. Additional layers, including Global Average Pooling, Dropout, and Dense layers, were added to perform multi-class classification. The experimental results showed that the proposed model achieved a test accuracy of 0.8893 with a test loss of 0.5587. Furthermore, the macro-average metrics obtained from the classification report were 0.9018 for precision, 0.8893 for recall, and 0.8877 for F1-score. These results indicate that the model was able to classify most activity classes effectively. However, several classes with similar visual characteristics still caused misclassification, indicating opportunities for further improvement in human posture recognition performance.