This paper presents an integrated framework for real-time target category recognition integrating the active template matching and the context of visual surveillance. The active templates are in the form of a dictionary of active features (bases), which are allowed to slightly shift at different locations and orientations. They can be learned for each object type from a small set of positive samples that roughly aligned. With these learned deformable templates, the moving foregrounds subtracted from background model are recognized through searching maximum matching likelihood. To avoid the exhaustive search for template matching and reduce the noise disturbance, a scheme to estimate target size and pose at specific location is developed based on the contextual information of scene geometry. This framework can be an independent module embedded into a visual surveillance system. Its performance and benefit of using context are quantitatively demonstrated on public dataset with comparisons.
展开▼
