Carried object detection in short video sequences

In this paper we propose an approach to carried object detection in short video sequences and overcome some of the shortcomings of other methods. The proposed approach can estimate pedestrian´s walking direction accurately without manual calibration and classify the probable carried object pixels with no need for prior information. Pedestrian´s silhouette is first aligned to get the temporal template. Then, the temporal template is matched against a set of exemplars using Principle Component Analysis (PCA) and exhaustive search. Finally, a fuzzy cluster method is applied to classify the protruding pixels. The method has been tested on PETS 2006 dataset and the detection of carried object is accurate and robust.