Introduction

This paper presents recent work we have on image understanding in the context of visual surveillance applications ¹. We propose a video understanding framework based on a priori knowledge. This work is based on three hypotheses: first we consider a static camera, second we use a unique monocular camera and third we deal with real-time constraints. The first hypothesis (static camera) is often verified for current visual surveillance networks and allows us to simplify the low-level detection of mobile objects w.r.t. a fixed environment. The second hypothesis (unique monocular camera) is verified in almost all current visual surveillance networks. The third hypothesis (real-time constraints) is very interesting as it implies that the solutions should be kept with a minimal computing time. But it also implies a fully automated system. First, after a presentation of related work, we introduce the general scheme of our approach which is based on the use of predefined scenarios and a priori contextual information. Second, we detail the current low-level image processing techniques used for mobile object detection and tracking. Third, we describe the role of a priori contextual information and different ways of representing this information. Then we adress the problem of high-level description of mobile object behavior using generic observable events and application-dependent scenarios. Finally, results obtained on different visual surveillance applications in the european Esprit project AVS-PV are shown and discussed. The paper concludes with future work for enhancing the robustness of such image understanding systems and for improving their capabilities of re-use.