Video-based People Counting System Algorithm Design

Pedestrian counting plays an important role in many industries. It helps people use limited resources sensibly and efficiently, and helps people monitor the activities of crowd in a larger building (such as sports venues, exhibition hall) to prevent overcrowding and ensure the safety of the people. It also can help people evaluate whether the facilities (such as seats, public health infrastructure) in buildings meet the customers’demand, whether the facilities in buildings is installed conveniently and reasonably. So pedestrian counting is very valuable. With the development of computer technology and image processing technology, video-based intelligent monitoring system has been widely used. This Create the conditions for the emergence of video-based Pedestrian counting system. This thesis designed a video-based Pedestrian counting system.Pedestrian counting system designed in this thesis is mainly used in entrance and exit. The theory of digital image processing and pattern recognition methods is widely used in this system. The main work of this thesis:1. A new Gaussian mixture model algorithm was proposed. This new Gaussian mixture model was combined with spatial information. When new Gaussians mixture model was built for each pixel, these pixels which near the center pixel were considered. Dynamic weights learning rate was applied to adapt to environmental changes as soon as possible. The results of experiment showed that this new Gaussian mixture model can adapt to the dynamic background.2. Negative samples are always limited in the machine learning algorithm, but the practical application environment is complex and diverse. So the classifier which is trained offline may not adapt to all situations. To solve this problem, the positive samples and the limited negative samples related to practical background were used to train a new classifier in this thesis.3. The traditional mean-shift algorithm usually use color histogram to describe the object, but traditional color histogram don’t combine spatial information. It easily causes losing targets which have similar color distribution. To solve this problem, this thesis proposes a novel mean-shift algorithm based on block color histogram. The results of experiment showed that this algorithm had a good performance when tracking similar target.The results of experiment showed that the calculation accuracy of the pedestrian counting system designed in this thesis can reach95%, when pedestrians carry objects and are moderately crowded.