The Design Of People Counting Systembased On Vision

Passenger quantity is the statistical data of people through the specified area, it has great importance in terms of business, security, resource scheduling. For example, with the statistics of passenger quantity in different periods and regions in malls, you can adjust the marketing strategy and store location to increasing economic efficiency. With the statistics of passenger quantity in public places, such as stadium, security personnel and facilities can be reasonably allocated to strengthen the response capacity of emergency. With the development of computer technology, image processing and pattern recognition technologies, using computer to do the image analysis and count people become possible. People counting based on vision can overcome shortcomings, such as imprecision and installation restrictions, of the past methods. It attracts more and more scholars.This article divides the people counting system based on vision into three sections: a region of interest extraction, pedestrian detection, pedestrian tracking. Each of these three sections is studied:1. Rectangular motion region extraction. In this paper, the motion region is extracted as a region of interest for subsequent detection. The extraction reduces the computation of subsequent detection, removes the interference of complex background and improves the detection accuracy. This paper uses vibe algorithm to extract the motion region, mathematical morphology to filter the extraction results, and external rectangle to get final results, which will be treated as a region of interest. Vibe has simple computing, high calculation speed, and robustness to the changes of background. Mathematical morphology filter can remove the small areas and holes in extraction results and improve the integrity and accuracy of extraction.2. Pedestrian detection based on head-shoulder feature. This paper detect pedestrian through head-shoulder feature and it helps to overcome the lacks of the past pedestrian detection in the scene of people counting. Under the people counting scenes, pedestrian detection through the whole body gets poor performance because of the occlusion between pedestrians. People’s heads vary greatly due to the hair, decoration and other reasons, so head detection also can’t get good detection. Head-shoulder has a "Ω" shape and contains a wealth of outline information. From the plan view, the occlusion is not serious even with a large pedestrian density. This paper chooses CENTRIST feature with its characteristic of encoding the contour information of target, and use the support vector machine to train a classifier for pedestrian detection through head-shoulder features.3. Multi-target tracking based on Kalman filter. Kalman filter is suitable for computer implementation because of its recursive form and it’s fit for real-time systems due to its high calculation speed and small storage capacity. Hungarian algorithm is a simple solution of assignment problem. This paper uses Kalman filter to get predictions of detection results, treats the correspondence between the predictions and detections of next frame as assignment problem and uses the Hungarian algorithm to solve it. The Euclidean distance between predictions and detections is composed of the cost matrix.In this paper, rectangular motion region extraction, pedestrian detection based on head-shoulder feature and multi-target tracking based on Kalman filter are applied to the people counting system. Experiments show that this system can achieve real-time, accurate statistical results.