Top View Person Detection and Counting for Low Compute Embedded Platforms
Prashant Maheshwari, Doney Alex, Sumandeep Banerjee, Saurav Behera, Subrat Panda · 2018
In this paper, we present an optimised approach for the top-view person detection and counting for low compute embedded platform. Earlier methods have used background subtraction for top-view detection which produces inaccurate results due to merging of blobs when there are multiple people in the frame. Several Deep Learning methods have been proposed for detection and tracking but they are computationally expensive to run on low compute device. We present an adaboost classifier for top-view detection and Kalman filter-Hungarian assignment for tracking which is optimised to give high accuracy on a low compute embedded platform. We achieved a counting accuracy of 97% in real-time (upto 40 FPS) on a Raspberry Pi3B. We also present a heuristic approach to handle false positives in real time through dynamic learning and unlearning of detections along with other optimisations in tracking.