Research On People Counting Methods Based On RGB-D

At present,the applications of computer vision technology in the field of video surveillance have gradually increased.How to count the number of people rapidly and accurately by adopting image processing technology has become a key research task.The common color(RGB)images which include rich texture and color information are easily affected by many factors,such as shadow,illumination changing and so on.The depth(Depth)images can work stable under these situations,whereas it is lack of available texture information.According to aforementioned situations,this thesis mainly research on people counting methods based on RGB-D images.The specific work consists of the following three aspects:The first part of this thesis research on people counting methods based on color and depth bi-module data.The purpose of this task is to count the current number of people within a given area.Aiming at this problem,this thesis firstly segment the human objects from the depth images rapidly by adopting the image segmentation method.Then,CNN model is used to identify the candidates in color images.Therefore,the task of people counting is completed in given areas.Experimental results show that the segmentation step in this method can segment the human objects from depth images effectively,and the objects recognition step can recognize the real objects in candidates correctly.The second part of this thesis focuses on the issue of people-flow counting in complex environments.Different from the task above,people-flow counting means to count the number of pedestrian through the surveillance areas in a period of time.The techniques adopted in this task need to detect and track the pedestrian objects simultaneously.For this purpose,the Kinect sensor used to capture images in this part is installed in top-down view.Firstly,the method proposed in this thesis adopts a background subtraction technique to fast obtain the moving regions on depth images.Secondly,the water filling algorithm is used to effectively detect head candidates on the moving regions.Thirdly,a SVM model is applied to recognize the real heads from candidates.Finally,this thesis adopts a weighted K Nearest Neighbor based multi-target tracking method to track each confirmed head and count the people through the surveillance area.Four datasets constructed from two surveillance scenes are used to evaluate the proposed method.Experimental results show that our method outperform the state-of-the-art methods and our method can work stably on condition of kinds of interruptions.In the third part of this thesis,based on the people-flow counting algorithm,a Kinect based people-flow counting software system is designed and implemented.The system is programmed by C++ and OpenCV,which can count the number of pedestrian through the surveillance areas in real time.