Study Of People-counting System Based On Kinect

Research on people counting system has been being a very active field of machinevision. It has a wide range of applications to count and analysis the flow of people,especially in large public spaces such as supermarkets, shopping malls, andsubways.Maximal social and economic benefits can be obtained by analyzing thefeedback data from people counting system and allocating resources reasonably andeffectively. Automatic people counting techniques are involved with a wide range ofknowledge such as image processing, machine vision and pattern recognition.Although there are many works and some mature algorithms on people countingsystem, it still is difficult to be robust against the variations of background and light,and occlusions among objects. Objects are three-dimensional in the world. However,traditional people counting system which just use two-dimensional information havedifficulties in solving these problems mentioned before. Moreover, traditional algori-thms which reconstruct three-dimensional information of objects are computationallyexpensive and need the support of equipment with high costs; it can’t meet with therequirements of real-time processing and general applications. Launched by Microsoftin2010, at a low price, Kinect depth sensor can capture color images and depth imagessynchronously and facilitate research towards people counting system.This paper is mainly on counting people in and out of large public spaces with thesupport of Kinect depth sensor and other hardware devices. A software framework isdesigned, and it includes three main parts, i.e., pedestrian target detection, targettracking and pedestrian counting. In this paper, noises are removed from depth imagesat pre-processing stage using morphological processing operations and an improvedmedian filtering algorithm.For detection of objects, the Maximally Stable Extreme Regions algorithm(MSER) is discussed. Considering the deficiency of its accuracy and speed, an effici-ent improvement is proposed. Firstly, the depth images are divided into nonover-lapping regions, and then MSERs run in all regions simultaneously. At last, humanheads can be obtained by applying the head constraints to the stable regions and thenlabeled in the color image synchronous acquired through the corresponding relation-ships. Experimental results show that our improved segmentation algorithm achieveshigher accuracy and run faster. For tracking and counting of pedestrians, Kalman filter algorithm is employed topre-estimate the positions of pedestrians, then pedestrians are tracked using regionaltemplate tracking algorithm combined with centroid algorithm when they enter thedetection area. Moreover, we present a two-way people counting strategy that pedes-trians are counted only when the centroids of pedestrians cross the two counting lines.Finally, people counting system is programmed according to the algorithm out-lined in this paper. The feasibility of the system is verified using images captured inour laboratory, and the results show that the system is robust and can achieve a highaccuracy.