A method for multiple rigid-object motion segmentation based on detection and consistent matching of relevant points in image sequences

This article describes a method to detect and characterise the independent movement of several rigid objects in a sequence of images. No model is assumed for the moving objects, meanwhile a linear model in the image plane is accepted for the pixels of each object, which can be changed without losing the generality. The approach works on each pair of images. Relevant points are detected on them and then each moving object is identified as two clusters of similar relevant points, one per image, that perform a consistent movement. The consistent movement constraint is incorporated into the cluster building phase, which is different from previous approaches. The main property of the proposed method is its robustness, since minimal assumptions are assumed. Results are provided with tests on two different kinds of scenes: a laboratory one and a natural traffic scene