Pedestrian Identification, Tracking and Counting in Video Images

• Main Authors: Yi-Fan Wu, 吳一凡
• Other Authors: Chi-Cheng Cheng
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/56699643648943045220

碩士 === 國立中山大學 === 機械與機電工程學系研究所 === 104 === Abstract Pedestrian counting is a way to apply object detection and tracking technology to count the number of pedestrians who enter the area of interest for a period of time. According to the head-body characteristics of pedestrians, this thesis proposes a simple and feasible method for pedestrian tracking based on the BLOB (Binary Large Objects) matching approach, which can achieve the tracking mission by labeling every target and effectively solve the problem of pedestrian occlusion. Firstly, the median filter is employed to remove possible noises, and background is extracted by the improved mixed Gaussian model. Combining the mixed Gaussian model and the background subtraction shows better performance and adaptability compared to the traditional Gaussian model approach. After the moving objects are detected by image preprocessing, the pedestrian can be identified by the HOG (Histogram of Oriented Gradient) features and the SVM (Support Vector Machine) classifier. In order to predict pedestrian’s trajectory, the Kalman filter with the BLOB method are chosen to improve computational efficiency by narrowing the searching region. Tracking pedestrians in the pre-assigned target area is able to reduce misjudgment of objects caused by overlapping. Two-way counting can also be accomplished via pedestrians crossing a given counting line. The person datasets in experimental verification contain 1500 positive samples and 12000 negative samples. 420 hard examples, which bring about wrong discriminate results for the initial classifier, are also added into the negative samples to enhance classification capability. The experimental results on identification and counting of pedestrians for the test video demonstrate 90% successful recognition rate and 60 ms average processing time. In the actual video through the continuous training of the classifier, the final successful recognition rate can reach 82% and the average processing time becomes 120 ms. Although some misjudgments still exist, the missing rate is only 10%, which should be in the acceptable range by taking into account uncertainty in actual shooting environment. The presented method in this thesis can effectively provide function of people counting. The irregular target area and the counting line can be set as the user’s wish. This flexibility will be helpful for different environments and applications in the future.