A Top-View People Counting System Based On Kinect And Jetson TK1

The people counting system plays a significant role in the public security area,the comprehensive arrangement and the resources allocation.However,the people counting system based on computer vision techniques exhibits a range of limitations in practical tasks.On the one hand,most of the vision based people counting algorithms needs to be implemented on the PC,which is bulky and not portable,resulting in the difficulty of system installation and maintenance.On the other hand,the RGB image based people counting system is variant with light,texture,occlusion and other transformations,which lead to limited accuracy and robustness.In this paper,the advances of 3D sensor technology is combined with embedded technology for designing a real-time statistics and portable embedded people counting system,The whole system is implemented with the second generation Kinect and embedded development platform Jetson TK1 which consists of three parts.1.The construction of hardware platform.Based on the comprehensive survey of widely applied somatosensory camera and embedded development platform,this paper firstly proposed a hardware platform of people counting system by using somatosensory camera Kinect and embedded computing device Jetson TK1.Additionally,this paper demonstarte a communication mechanism among those two parts,which enables the data transmission from Kinect to Jetson TK1.2.The design and implementation of people counting.The overall people counting algorithm includes four stages: depth image preprocessing,foreground extraction,head segmentation and recognition,tracking and counting.Preprocessing stage recover the missing pixels of depth image,to ensure the smoothness in subsequent processing.After obtaining the recovered depth image,Gaussian Mixture Model is adopted to construct and update the background model,followed by moving people extraction by using background subtraction method.Considering the difficulty in depth image based people detection which is caused by the difference of the crowd's height and head size,as well as the occlusion problem,in this paper,the robust,locality and invariant water filling algorithm is exploited for head segmentation and detection.Due to the lack of color and texture information of depth image,the RGB image based tracking algorithm is not suitable in this framework.In addition,most of the RGB image based algorithms are complex,and thus not suitable for embedded computing with limited resources.In this paper,an efficient depth image based centroid tracking algorithm is proposed.Tracking people are performed by measuring the distance of head center point before and after the frame.After tracking a number of frames,the direction of entry and count are obtained.3.The design and implementation of mobile display terminal.In order to directly obtain the real-time statistics of the number of people,this paper designs a mobile display terminal by employing Zigbee wireless data transparent transmission module.The display terminal can receive and demonstrate the count result of the extended Zigbee module of the Jetson TK1 by wireless communication with in a certain distance.In the end,the comparative experiments and the system experiments are conducted to validate the robustness and accuracy of the proposed system.The experimental results illustrate that the embedded system achieves the merits of high accuracy,real-time performance and portable,which meets the requirements of different practical applications.