Summarizing Nursery School Surveillance Videos by Distance Metric Learning

Abstract

We propose a novel video summarization approach that takes the mass quantity of nursery school surveillance videos as input and produces short daily video digests for children. The proposed approach makes full use of a distance metric, which is learned using a novel learning algorithm called the adaptive large margin nearest neighbor (ALMNN), and can properly measure the similarity between video clips. The learned distance metric is combined with supervised classification and unsupervised clustering to categorize daily raw surveillance videos into individual event categories. The final digest is constructed by selecting representative video clips that belong to individual event categories. Digests generated using our approach cover and reflect the various activities of children in nursery schools. They are of interest to parents, and they also enable easy access to mass quantities of daily surveillance video data. We implemented the approach as a practical system in a real nursery school environment and assessed its performance.