Accurate activity recognition in smart environments is vital for assessing an occupant’s ability to perform Activities of Daily Living (ADLs), indicating independent living. However, sensor errors like false readings and missing data often reduce system reliability. Traditional methods - Neural Networks, Dynamic Bayesian Networks, and Hidden Markov Models struggle with uncertainty in incomplete or unreliable sensor data. This work applies Dempster-Shafer (DS) theory to better handle such uncertainty, leveraging its reasoning mechanism to manage sensor inaccuracies. Using a dataset of nine representative ADLs from 32 activities captured via first-person video, with real-world imperfections like artifacts and frame loss, the DS-based system was evaluated against a traditional machine learning approach. Results show DS theory improves recognition performance under high uncertainty, demonstrating its potential for real-world smart environment applications.