Research Of Pedestrian Flow Statistics Based On Video Images

Automatic statistics of pedestrian flow is a very important content in intelligent video monitor. It has been used in the fields of traffic management, business decision-making, public safety monitoring and so on. Currently, statistics of pedestrian flow mainly relies on manual statistics. Pedestrian statistics has two problems when pedestrians are sheltered, one is missing detection and the other is error statistics. The disadvantage of this method not only wastes a large amount of time, but also counts the result inaccurately. In view of the existing problems in the system of pedestrian flow statistics, an effective system is designed and implemented about automatic statistics of pedestrian flow in this thesis. The system has two main sections, one is pedestrian detection and the other one is pedestrian statistics.In the part of pedestrian detection, a head detection method based on Harr feature classifier is studied and implemented. The approach is detecting head target by a head cascade classifier model in video monitor, and the model is obtained by training head images of positive and negative samples with Adaboost, so as to reduce missing detection effectively when pedestrians are sheltered.In the part of pedestrian target statistics, a matching method is proposed, the approach is Hu invariant moments combined with Euclidean distance. The proposed approach of counting the pedestrian flow has two sections. Firstly, tracking objects with the Camshift algorithm that is forecasted by Kalman; secondly, determining whether the objects belong to the same goal or not, the determining condition is Hu invariant moments and Euclidean distance of adjacent frame object, so as to reduce error statistics of pedestrian target effectively.Firstly, automatic statistics system of pedestrian flow based on OpenCV is implemented in the Windows operating system, furthermore, it is tested and analyzed. The results of experiment show that the average accuracy is 98.2% with the proposed statistical method, and the average time of the statistic for each pedestrian target is 19 ms. In order to improve the portability of the system, the designed system is transplanted to Android development board based on Linux successfully.