An Ellipse Fitted Training-Less Model for Pedestrian Detection
Arindam Sikdar, Ananda Shankar Chowdhury · 2017
The problem of pedestrian detection has gained much popularity in the computer vision community in recent times. We have noted that the existing solutions to this problem are mostly supervised in nature. However, it is difficult to guarantee availability of labelled training data in all situations. In this paper, we propose a training-less solution of pedestrian detection. Some of the additional challenges for pedestrian detection are proper handling of viewpoint dependencies, background clutter, illumination variation and occlusion. We design an ellipse fitting model, as a part of our training-less solution, for accurate pedestrian detection. In this model, we fit an ellipse to each competing bounding box (proposal). An area and entropy based quality factor is introduced for every such (fitted) ellipse to discriminate among the proposals. We filter out proposals with low quality factors. Performance comparisons with some well-known supervised pedestrian detection approaches on publicly available PETS2009 dataset demonstrate that our solution is highly promising.