Human activity recognition plays a crucial role in fields such as healthcare and security, where accurately identifying actions and transitions significantly enhances the effectiveness of monitoring systems. This study evaluates the performance of deep learning architectures including convolutional neural networks (CNN), long short-term memory networks (LSTM), gated recurrent units (GRU), and hybrid models such as CNN-GRU and CNN-LSTM using sensor data from smartphones. Time series were analyzed in both the time and frequency domains to classify dynamic activities and postural transitions. Results show that analysis in the time domain consistently yielded superior performance, with the CNN-GRU model achieving 97% accuracy, outperforming other architectures in both categories. Despite challenges due to class imbalance, hybrid models proved most effective, highlighting the importance of combining spatial and temporal analysis. These findings reinforce the potential of hybrid architectures in addressing human activity recognition complexity in real-world applications.