| Pedestrians counting methods are widely used in transportation, business, etc. The development and maturity of computer vision technology makes it become a hot research topic. Pedestrians counting methods based on computer vision technology have many advantages, such as its convenience to obtain the scene, wide detection range, easy to integrate with monitoring system. In view of the crowding scene, we propose a real-time pedestrians counting algorithm based on Haar and HOG. Our research mainly focus on the accuracy and the real-time performance of pedestrians counting. The main content of the dissertation are as follows:1. We design a head detection structure which based on Haar features and HOG features. (1) As the fast detection speed of the Haar classifier, we use it to obtain the candidate areas of human heads. (2) As the high detection accuracy of HOG classifier, we use it to filter the candidate head areas to find the final heads. Through this method, we can solve the problem of the slow detection speed of HOG classifier, meeting the real-time performance of the whole system. At the same time, the fusion of Haar and HOG reduce the error detection rate to ensure the reliability of the pedestrians counting.2. The structure of the traditional Adaboost algorithm is modified to improve the performance of the Haar classifier. In the training process of improved Adaboost, we furtherly divide those error-classified samples into two classes:the missing detection and the false detection, and assign more weight values on the missing detection. This discriminatory weight assign strategy leads to lower omission rate.3. We propose a spatio-temporal correlation analysis algorithm to complete fast pedestrians tracking and counting task. We make use of the clustering thought to create a human head model for a detection target. Through the color, shape and some other characteristics of the head model, we can accomplish the tracking and counting of pedestrians. |
PDF Full Text Request