Person detector for different overhead views using machine learning

We explore a dimension of detecting people with a completely different perspective i.e. use of a top view. An overhead view is often preferred in the cluttered environments because looking down from a top view can afford better coverage and much visibility of a scene. However human detection in such or any other such type of extreme view can be challenging. The reason is that depending on the positions of people in the picture or image, there can be a significant variations in the poses and appearances of a person. To handle all such variety of poses, appearances and body articulations from the perspective of a top view, we propose a novel technique which transforms the region of interest containing a human to standardized the shape. After that applying Rotated Histogram of Oriented Gradient (RHOG) algorithm with machine learning based SVM classifier improves detection performance significantly. We show the potential of our proposed RHOG algorithm across different scenes. When a classifier trained on SCOVIS dataset and applied to our newly recorded overhead datasets named SOTON and IMS, respectively. We achieve a detection rate of 96% and 94%, respectively.