The Study on Counting Method for the Moving Crowd Based on Depth and Color Information

• Main Authors: Chih-Pin Du, 杜志斌
• Other Authors: Chao-Ho Chen
• Format: Others
• Language: zh-TW
• Published: 101
• Online Access: http://ndltd.ncl.edu.tw/handle/60467758322002334390

碩士 === 國立高雄應用科技大學 === 電子工程系 === 100 === Currently, there are many literatures of people counter was based on video processing. The most studies used 2-D color image to be principally image data, and used well-known technology (e.g., Background Subtraction, Frame Difference …) to do people detect and segment. They have well-count at the sparse people flow, but its accuracy muse be reduce when they are used in crowd. The common research of people counter are used for the number of the moving object pixel, moving object’s height or width, horizontal project, vertical project to estimate the number of people. Although it can estimate the number of people flow, it can’t keep high accuracy. For all the reasons above, we propose a new method to combine the 2-D color image and the 3-D depth image to do the crowd detection, segment, and raise the accuracy of crowd counting. This paper presents an automatic crowd counting system, and it is based on the Kinect to capture the color and depth image to do crowd detection, segment and count. We set up the Kinect at the entrance and let it vertically downward to capture the vertical view. We use the vertical view image to get more precise crowd counting. This paper is based on the depth image to segment the foreground object and then use the color image to get more precise image segment. On the technology of people detect and segmenting. First of all, we are setting a value to filter the original depth image and then we can extract the object by using the technology of MER. Second, we use like the watershed technology to do accurately segment and then we can get the high value region of the image, so that we can catch the whole people information. Last, we use the new algorithm named Region Segment to extract head, shoulder, etc. After that we use the histogram and color feature to judge the object is people or not. If the object is the people, we track it and counting. Finally, we use our method to count people with the video, which we film the crowd video. The system has more than 95% accuracy.