Fast binary shape categorization

摘要

A novel approach for object categorization suitable for video surveillance is proposed. We describe shapes only using radius and arclength of their curvatures, which allow differentiating between objects that appear in the monitored area. We conducted experiments on classes such as pedestrians, cars, cyclists, and animals (horse, cow, dog, and cat). Our approach achieves a reasonable accuracy (95.66%) on Kimia dataset, surpasses the accuracy of the state-of-the-art methods (93.75%) on CDnet videos, and allows handling cases of object merge and split usually present in foreground masks issued from background subtraction of videos.