Indoor Real-time People Counting Algorithms And Design

People counting techniques have been applied in many public places with entrances, such as supermarkets, subways, bus stations and so on. The people flow data of these scenes can supply useful information for public security, marketing decision and resource allocation. With the rapid development of image processing techniques, current people counting methods are becoming more and more intelligent and automatic than before. But the current methods still have many problems, such as inaccurate statistics results, poor robustness and so on. With the increasing requirements for automatic people counting systems based on digital image processing and computer vision, effective people counting methods become remarkable and meaningful.This paper proposes two kinds of efficient and accurate people counting algorithms and schemes according to the problems above: one is based on head detection and tracking to calculate the number of people who move under an over-head camera, namely perpendicular people counting, and the other is based on head-shoulder detection to evaluate people flow in the monitoring area of an oblique-style camera, namely regional people counting. There are four main parts in the first kind of people counting: object region extraction, head detection, head tracking, and crossing-line judgment. The extraction of object region utilizes an effective method to obtain foreground regions including moving people, and uses some post-processing methods to optimize results. The stage of head detection exploits an off-line trained Adaboost classifier by LBP feature for head detection. Head tracking stage uses a kind of local head tracking algorithm which is based on Meanshift iteration to track head objects. Combined with counting line and region, the crossing-line judgment algorithm determines whether the candidate head object will be counted or not. As to the second kind of people counting, there are mainly steps of head-shoulder detection, and it uses number of human objects in head-shoulder detection to estimate crowd density of the monitoring region. The head-shoulder detection algorithm employs a classifier which cascades a primary Adaboost classifier and a secondary SVM classifier, to implement rough detection and precise filtering of head-shoulder objects.The perpendicular people counting method proposed can achieve promising people counting accuracy and acceptable computation speed, and is suitable for the real-time applications. In addition, the regional people counting method proposed can assist the former method to implement combination of two types of people counting methods. The proposed method works well in different scenes and various crowd densities. Experiment data shows that our method can obtain promising people counting accuracy about 96% and acceptable computation speed about 30 frames per second. Compared with current methods, the propos ed method has a promotion in accuracy, real-time performance and robustness.