In the development of the next generation of smart homes and remote health monitoring system, human action analyzing algorithms are of vital importance. Among different sensor technologies, vision-based systems are superior in the sense that they can provide a non-intrusive interface between human occupants and the environment. It is almost impossible to build an efficient system for human action recognition without fine-tuning and evaluating its performance on a realistic dataset. An important challenge here is the absence of such a comprehensive dataset. To address this issue, we introduce a new dataset designed for human action recognition applications in a smart home environment. The performance of the existing human action recognition algorithms is tested using this dataset. In addition, we propose a heuristic approach based on error-correction codes to prioritized different actions in the learning process and improve the recognition accuracy for difficult actions up to 17%.