HAR in uncontrolled environments presents significant challenges due to variability in sensor data, noise, and dynamic user behaviours. This chapter systematically evaluates the performance of shallow, ensemble, and deep learning models in addressing these challenges using three datasets: a custom real-world dataset, MotionSense, and mHealth. Shallow models (decision tree, K-nearest neighbours, Naive Bayes) demonstrate computational efficiency but struggle with complex spatiotemporal patterns, achieving modest accuracies (68–97%). Ensemble methods (random forest, XGBoost) improve robustness, attaining 96–99% accuracy by aggregating diverse base learners, albeit with higher computational costs. Deep learning models (ANN, LSTM, CNN-LSTM) excel in automatic feature extraction, with CNN-LSTM achieving peak accuracies of 98.89% by leveraging spatial and temporal dependencies in sensor data. However, deep models demand substantial computational resources, with training times up to 2634 s, highlighting a trade-off between accuracy and efficiency. The study underscores CNN-LSTM’s superiority in uncontrolled settings but emphasizes ensemble methods as a pragmatic choice for resource-constrained deployments. These insights guide the design of scalable, accurate HAR systems, balancing performance and practicality. Future work will focus on lightweight hybrid architectures and edge-compatible optimizations for real-world applications.