Transfer Learning Approach for Human Activity Recognition using EfficientNet
R. R. Rajalaxmi, R ThamilSelvan, P P Sudharsana, S Ponharishkumar, S Ragavi, Sanjeev Rathan R · 2025
Human walking activity recognition is concerned with correctly identifying actions carried out by the users, whereas gait analysis identifies abnormalities in such actions. Recognizing activities from color images presents some of the challenges, like variation in background, lighting, camera movement, and differences in appearance. A sensor-free solution addressing these issues is much required, especially in the health sector. The studies on deep learning proposed many approaches and methods, such as an ensemble system processing walking activity image. In this study, deep learning based human activity classification is performed without any need for sensor utilization. Preprocessing steps include resizing, median filtering, and contrast stretching. The skeletal representations are generated using MoveNet and MediaPipe. The EfficientNet B0 model is used for feature extraction, and YOLO V8 is for improving the detection of body parts. The key points obtained are further grouped and organized using an MLP. This approach reached an accuracy of 43.81% using a YOLO V8 model with the classification layer of EfficientNet on a pre-processed dataset. This framework shows great promise for strong real-time recognition of human actions in the healthcare and industrial domains.