Аннотация
The aim of this paper is to present the algorithms that were developed for detecting human in the conditions of overlapping and non-overlapping. Overlapping means when person is not fully visible in the image and occluded by another person at the front or right/left side. In order to achieve this goal, three steps were implemented. The algorithms were implemented in C++ with the help of Open Source Computer Vision (OpenCV) library. The first step was human detection when single person's body occupied most of the image. For this, three machine learning techniques were examined: K - Nearest Neighbour algorithm (KNN), Support Vector Machine (SVM), and Decision Tree learning. Their performances were compared and optimal classifier was chosen. The results of experiments have shown that SVM, in comparison with others, had the highest accuracy rate. It exceeded others for about 10-12%. So, SVM was taken as a classifier which detected human at an accuracy rate of about 95%. The second step was implementation of sliding window approach to detect multiple people in the image. And the last, third main step was people detection in conditions of occlusion. As it's known, SVM gives the decision value (confidence) for detection, negative value is for person and positive value is for non-person. Our approach was based on the assumption that if it is occluded person which overlap with detected person, its confidence should be smaller than confidences of its neighbour windows so that occluded person would be more `human like' with respect to neighbours. So, we found local minima within the windows, which intersect with a window of detected for person more than 45%. The results of tests for detection of occluded people indicated the accuracy rate of about 54% is electronic document is a "live" template and already defines the components of your paper.