Motion Tracking with Normalized Cut and Minimum Spanning Tree

We propose using the Normalized Cut method for motion tracking (J. Shi and J. Malik . “Motion Segmentation and Tracking Using Normalized Cuts”. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8), pages 888 905, 2000) on regions in frames that were identified by a Minimum Spanning Tree (MST) method in a pre-proccessing step. The purpose of the pre-proccessing step is to reduce the spatial resolution without losing important image information. An energy function, based on some selected properties of regions, is then calculated for each pair of regions in a fixed number of consecutive frames. This energy function represents the similarity of two regions. Based on this similarity, the Normalized Cut method is used to identify salient groups of regions. Finally, corresponding salient groups of regions of neighbouring sets of consecutive frames are found. We show on different experiments, how pre-segmentation can help to reduce computation time by reducing the spatial resolution of the input frames. We tested different methods to reduce temporal resolution to gain an additional speedup.