Optimisation of Feature Space for People Detection from TopView on Light Embedded Platform
Doney Alex, Prashant Maheshwari, Sumandeep Banerjee, Subrat Panda · 2018
Detecting people from a top view video feed is comparatively a more difficult problem than pedestrian detection especially on a real time system, as the perceived 2D shape of person from the overhead view is of widely varying nature due to perspective distortions with almost no unique discernible outline shape. Therefore the amount of information available from overhead view is very less. Also, there is a lot of rotational variation of the data from top view. However the advantage of having such a detector is that for a tracking application there would be little occlusion. Hence, this problem is worthy of specialised attention. In recent times there are many deep learning based approaches, but all of them are computationally expensive. We present a method to effectively train a computationally light AdaBoost classifier based detector, which uses the limited amount of information and can give a high accuracy running on a light embedded platform such as Raspberry Pi 3B.