Multi-Saliency Map and Machine Learning Based Human Detection for the Embedded Top-View Imaging System

Abstract

Compared to the side view, a top-view is robust against occlusion generated by objects located indoors. It offers a better wide view angle and much visibility of a scene. However, there are still problems to be handled. The top-view image shows asymmetrical features and radially distorted scenes around the corners, such as omnidirectional view images and self-occlusion. Conventional human detection methods are suitable for finding moving objects in front view imaging systems. And there are some limitations, such as slow execution speed due to computational complexity. In this paper, we propose an efficient method. A static saliency map with low activity and a dynamic saliency map with a lot of movement are respectively detected. These two models were fused to create a multi-saliency map, and both characteristics were used simultaneously to improve detection rates. To handle problems such as asymmetry, a rotation matrix was calculated around the center, and Histogram of Oriented Gradient (HOG) features descriptor were extracted from the multi-saliency map to create an image patch (a small image region of interest containing human candidates). For the classification of image patches, we used machine learning-based supervised learning models support-vector machine (SVM) algorithm to improve performance. As a result of the proposed algorithm, it showed low resource occupancy and achieved Average Precision of 92.3% and 96.12% when Intersection over Union were 50% and 45% respectively.