Online surveillance object classification with training data updating

Abstract

One of the main problems of object online classification using classifier trained offline is the mismatch of online test images and offline training data. In this paper, we propose an online video object classification algorithm with the mechanism of training data updating. By selecting part of the uncertain test data captured online and labeling them artificially to replace a proportion of the training data, the classifier can be retrained using the renew online training data, and thus the possible mismatch problem can be avoided and then higher classification accuracy can be achieved. From the experiments based on online surveillance video object classification, it was observed that: compared with existing classifier without training-data-updating, the proposed method can achieve up to average 18% classification accuracy increasing.