In this paper we describe a real-time gesture recognition algorithm based on contour signature images, generated by collecting a set of contour signatures over time. Each contour signature is generated in each frame by computing the polar coordinates of each point of the contour in relation to the center of mass of that contour. The gesture recognition is reduced to the problem of finding the correspondent contour signature image from a set of previously learned image contour signatures. This is achieved using a Karhunen-Loeve (KL) transform that reduces the set of trained images to a low-dimensional eigenspace, making recognition easier.

This method is particularly suitable for the recognition of gestures that can be perfectly described by silhouettes. Since no shape or 3D model is required the proposed method can easily be adapted to recognize several other types of gestures not necessarily performed by humans. The proposed algorithm is able to extract target gesture motions and classify them at the same time.

Preliminary experimental results demonstrate robustness and real-time performance of the proposed algorithm.