2D-Object Tracking Based on Projection-Histograms

Abstract

Image-sequence analysis for real-time applications requires high quality and highly efficient algorithms for tracking as there is no time to do the costly object recognition each time a new image is captured. Tracking with projection histograms revealed amazing results compared with standard correlation methods. Trackers based on projection histograms performed 31% up to 211% better than the reference methods on a common test set. The new template-based method relying on projection histograms ( RPH ) is described and compared with two commonly known template based methods namely the normalized cross-correlation ( NCC ) and displaced-frame-distance ( DFD ) methods. The input to the system consists of live or recorded video data where filterbased preprocessing can be applied before tracking in order to enhance features such as edges, textures etc. A region of interest (ROI) is taken as a template for tracking. In subsequent images tracking exploits a Kalman-filtered local search in order to renew correspondence between the object template and the new object location. Comparative tests were performed with real-live image-sequences taken in underground stations. Tracking with projection histograms outperformed tracking by NCC and DFD on grey-level image-sequences as well as on edge-enhanced image-sequences. Even the worst chosen parameter set for tracking by the new RPH method resulted in better tracking as with the best ones for both NCC and DFD .