Activity recognition has been receiving significant attention from a variety of research areas such as human performance enhancement, health promotion, and human computer interaction. However, recognizing activities from accelerometer data still remains a challenging problem due to sensitivity to sampling rates, misalignment of data, and increased variability in activities among clinically relevant populations. In order to solve these issues, we adopt methods from functional analysis, which consider non-elastic rate variations in movement. The overall framework factors out temporal variability within activity classes, before leveraging robust machine learning pipelines for a given end-use. The proposed approach has been evaluated on 7 classes of everyday activities with 50 subjects. The results indicate that proposed approach achieves improved performance with the improvements observed in separating similar classes that differ in temporal rates, and also demonstrate higher robustness to change in window lengths. These results suggest that temporal alignment should be considered a core part of activity recognition pipelines.