Daily Routine Recognition through Activity Spotting

This paper explores the possibility of using low-level activity spotting for daily routine recognition. Using occurrence statistics of lowlevel activities and simple classifiers based on their statistics allows to train a discriminative classifier for daily routine activities such as working and commuting. Using a recently published data set we find that the number of required low-level activities is surprisingly low, thus, enabling efficient algorithms for daily routine recognition through low-level activity spotting. More specifically we employ the JointBoosting-framework using low-level activity spotters as weak classiers. By using certain lowlevel activities as support, we achieve an overall recall rate of over 90% and precision rate of over 88%. Tuning down the weak classifiers using only 2.61% of the original data still yields recall and precision rates of 80% and 83%.