People are becoming more and more interested in human activity recognition (HAR) because smart devices are getting better and better everyday. These devices hold great potential in various fields, such as security, healthcare, and sports. This paper thoroughly examines HAR techniques for extracting meaningful characteristics from continuous temporal data gathered by motion sensors in wearable devices. The analysis explores techniques based on handcrafted features and highlighting the significant impact of deep learning methods. Individual models focused on deep learning often need help with generalization and robustness, as they may become too specialized and sensitive to variations in the data. To address this issue, an ensemble-based approach is utilized by combining the strengths of multiple models to enhance overall performance. This study focuses on introducing and assessing a hybrid deep learning model called CNN-GRU. This model utilizes convolutional neural network (CNN) for spatial feature extraction and gated recurrent unit (GRU) for capturing temporal relationships. The proposed model exhibited exceptional performance in our experiments, achieving 94.84% and 97.38% accuracy on the UCI-HAR and WISDM datasets, respectively.