Action segmentation based on Bag-of-Visual-Words models

Abstract

In this paper, we propose an action segmentation system which can segment a video into a number of video shots according to the human action. At the same time, the proposed system can also recognize the action types of those segmented shots. Our system estimates the action number of the video by using the distribution of visual word vectors without any predefined information and automatically adjusts the window step size instead of manual selection. In the proposed approach, the bag of visual word (BOVW) together with self-growing and self-organized neural gas (SGONG) are used to construct the action word dictionary to recognize the video shots. The experimental results show that the performance of our segmentation method can perform well for the test videos.