| Scene understanding is a task of studying high-level knowledge in the field of computer vision. It has been widely used in navigation, surveillance, human-computer interaction and other fields. It is important for researches and applications. In this thesis, we mainly study the method of modeling and reasoning belonging relations in scenes. There are three aspects study in this thesis:(1) We propose belonging relations as a qualitative representation for scenes. Objects are subject to gravity. All objects in a scene are stable, because one object is subject to a force that is opposite to the direction of gravity. We use a hierarchical structure, spanning tree, to represent all belonging relations between objects in a scene. We can construct the spanning tree by observing the spatial relation features in space.(2) We propose a belonging relation recognition method, which uses human-object interaction constraints. Without self-change relations, all relation changes are caused by human interaction with objects. We can infer establishments and changes of belonging relations from the perspective of the interaction between human and objects. We define three types of human-object interactions: take-out, put-in, and irrelevant. We extract the context features between human actions and manipulated objects. Then we model the relations of human-object interactions and belonging relations changes. We adopt SVM algorithm to recognize the belonging relations in a scene.(3) A spatial-temporal joint inference method to infer belonging relations between objects in scenes is proposed. We design an energy function, which includes a data term and a smooth term. The data term models the energy of belonging relations in spatial space, and the smooth term models the energy of belonging relation changes in temporal space. We cast inference into an optimization problem and our goal is to solve the minimum energy problem via dynamic programming framework.We collect a belonging relation change video dataset. We test our proposed methods on this dataset. The experimental results show that our algorithms can effectively recognize belonging relations and belonging relation changes in scenes. |
