Incremental learning of Bag-of-Sequencelets Model for Human Activity Recognition

Abstract

Recognition of human activity in video is a fundamental task in computer vision, and has been intensively studied. In activity recognition, most work has focused on developing good features or classifiers to increase the precision of classification. However, because the number of training videos has increased drastically, efficiency of learning is as important as precision of classification. Therefore, in this dissertation we focus on improving the efficiency of learning.

The first topic of this dissertation is class-incremetnal learning. Because huge numbers of videos are produced daily, the number of training videos is increasing. Most existing methods for activity recognition are not flexible enough to adapt to this increase, because they must retrain their learning models whenever new training videos with new activity classes are added for learning. In this dissertation, a Bag-of-Sequencelets Class-Incremental Learning (BoSCIL) is proposed to deal with this problem.

We also propose the BoS sub-category group merging (BoSSGM) method for efficient learning of a BoS model. BoSSGM enables efficient learning of BoS model by dividing the whole dataset into several sub-category groups. Moreover, it also enables merging of multiple datasets and recognition models for each dataset.

We validate our model on the several activity datasets which include large number of activity classes and various type of activities such as sports activities, simple actions and daily living activities. In experiments, we show that BoSCIL and BoSSGM reduce the training time greatly compared to the baseline BoS, while maintaining the classification precision.