Moving Shadow Detection using a Combined Geometric and Color Classification Approach

Automatic detection of moving objects is a fundamental problem in computer vision. Motion analysis, object recognition, and video surveillance applications often depend on reliable segmentation of moving objects against a fixed background. Although shadows move in the scene with the objects that cast them, it is often important that only objects in motion, and not their shadows, are detected. For example, false positives of moving shadows that are generated by current approaches to motion segmentation can lead to erroneous object models that will ultimately impact recognition rates. We present an algorithm for the automatic detection of shadows being cast by objects in motion on a planar surface that utilizes the additional geometric information afforded by two views of the scene. Foreground pixels are first generated by an existing motion segmentation method in each of the two views. Moving pixels that have no parallax lie in the ground plane and are considered to be potential shadow pixels. A subset of these shadow pixels is used as input to construct an adaptive mixture color probability model. Once the shadow color model is acquired, it can then be used to efficiently detect shadows in successive frames in conjunction with the zero parallax constraint. The approach has been tested in several different video surveillance scenarios. Results demonstrate that approximately 94 percent of shadow pixels can be correctly classified with small false positive rates.