We incrementally learn the human activity models with the newly arriving instances using an ensemble of SVM classifiers. It can retain the already learned information and does not require the storage of previously seen examples.
We reduce the expensive manual labeling of the incoming instances from the video stream using active learning. We achieved similar performances comparing to the state-of-the-arts with less amount of manually labeled data.
We propose a framework to incrementally learn the context model of the activities and the object attributes that we represent using a CRF.