Abstract:
The objective of this research was two-fold: 1) to develop lightweight deep learning algorithms that facilitate efficient learning of human daily activities, and 2) to utilize the resulting algorithm model to contribute to the development of innovative solutions that support human care in various fields. The research focused on three main areas: data pre-processing methods, data generation methods, and model training methods for classifying human daily activities. To conduct the experiments, a publicly available dataset called SPARS9x was utilized. The research employed Colab Pro as a tool and utilized the Python language for development purposes. The dataset included six different models, namely VGG16, ResNet18, PyramidNet18, Inception-V3, Xception, and EfficientNet-B0. The experimental findings revealed that PyramidNet18 achieved the highest accuracy of 99.15% and an F1-score of 99.15% when tested on the SPARS9x dataset. Additionally, it was discovered that hybrid models deployed with convolutional neural networks not only provided outstanding results but also demonstrated computational efficiency