A Vision-based Detection System for Walking-cross Activities of Pedestrians

• Main Authors: Liang, Chih-Bin, 梁志彬
• Other Authors: Fan Chun-Hai
• Format: Others
• Language: zh-TW
• Published: 1997
• Online Access: http://ndltd.ncl.edu.tw/handle/18872406598533197589

碩士 === 淡江大學 === 土木工程學系 === 85 === In this thesis, we try to use image processing technique to count pedestrians which are stationary in the image and track their trajectories when they are walking. In the pedestrian counting system, since the pixels which pedestrian possesses in the image are directly proportional to number of pedestrians, so we use background differencing method to establish regression model and calculate all people in the image. On the other hand, we use interframe differencing method to establish another regression model which is used to calculate moving pedestrians in the image. The difference between total number of pedestrians and number of moving pedestrians at the same time is considered to be the number of stationary pedestrians in the image. About pedestrian''s trajectory detection, we adopt template matching method based on image feature invariants to track each pedestrian''s template in sequence images. After linking every coordinates, we obtain pedestrians'' trajectory lines. The detection system can output four important parameters including pedestrians'' number, trajectories, speeds, and directions. A case study was demonstrated. About stationary pedestrians, the absolute diffenence between automaticcounts and real counts for each image is almost within 2 persons, and the ratio of maximum absolute error to the real count decrease gradually when people increase. About trajectory, the use of template matching method have only an average error of 1 pixel in x-axis, 2 pixel in y-axis. The error of speed is about 9.43%.