Boosting with Temporal Consistent Learners: An Application to Human Activity Recognition