Comparative Analysis of Classical and Deep Learning Approaches for Human Activity Recognition Using Smartphone Inertial Sensors

This work presents a study on human activity recognition (HAR) using smartphone sensors such as accelerometers and gyroscopes. The UCI HAR dataset is used to recognize activities including walking, sitting, standing, climbing stairs and lying down. Both classical machine learning models and deep learning approaches are employed, with the linear support vector classifier (linear SVC) achieving the best accuracy of about 96.67%. This study also explains data preprocessing, the extraction of 561 features and methods for handling data imbalance. Several steps are suggested for faster real-time deployment, such as pruning and quantization. This work is helpful for health tracking and daily activity analysis, demonstrating that simple models can still perform well on real sensor data. Leveraging a comprehensive dataset from the UCI repository, which includes accelerometer and gyroscope data, this study aims to predict various human activities such as walking, climbing stairs, sitting, standing and lying down. This research compares the performance of classical machine learning models and deep learning approaches, specifically focusing on long short-term memory (LSTM) networks. Key challenges addressed include feature engineering, data preprocessing and model optimization. The results show that the linear SVC outperforms other models, achieving high accuracy in activity classification. Evaluation metrics such as accuracy, confusion matrices and classification reports are employed to assess model performance. The richness of the dataset, enhanced by video annotations, provides valuable insights, although limitations such as dataset imbalance and other constraints are acknowledged. This paper concludes by discussing future research directions, including real-time application possibilities and the incorporation of temporal dynamics. Overall, this study provides a robust analysis of HAR using machine learning and offers practical insights for researchers and practitioners in the field.