[full paper] |
Nathanaël A. Rota, Monique Thonnat
We propose here a new approach for video sequence interpretation based on declarative models of activities. The aim of the video sequence interpretation is to recognize incrementaly certain situations, like states of the scene, events and scenarios, in a video stream, in order to understand what happens in the scene. The input of the activity recognition is an {\it a priori} model of the scene and human tracked in it. The activity recognition is composed of two subproblems. First, end-users have to declare all the activities in a configuration phase. Secondly, the declared models must be automatically recognized. To solve the first problem, we propose a homogeneous declarative formalism to describe all the activities (states of the scene, events and scenarios). The activities are described by the conditions between the objects of the scene. To solve the second problem, we translate it into a constraints satisfaction problem. Then, we use a classical CSP algorithm to recognize the activities in video sequences. Finally, we present some results to show the robustness of the approach.
Keywords: Reasoning about Actions and Change, Knowledge-Based Systems, Knowledge Representation, Vision
Citation: Nathanaël A. Rota, Monique Thonnat: Activity Recognition from Video Sequences using Declarative Models. In W.Horn (ed.): ECAI2000, Proceedings of the 14th European Conference on Artificial Intelligence, IOS Press, Amsterdam, 2000, pp.673-677.