People counting on video streams

Recently, the use of computer systems has become widespread with the developing technology. Object detection is a popular topic in computer vision and machine learning. The interaction of people with computers has increased and computer based systems have affected the human life so pedestrian detection methods are gaining more importance every day. Designed human counting systems are devices used to determine the number of pedestrian that pass through a door or corridor. These systems are usually used at the entrance of a large building, at railway stations, at airports, or at stops of public transport. Thus, the total number of people in the monitored area can be recorded. Human counting systems are very important systems for conducting marketing research surveys and for determining the human density in a given time period. In this thesis, an algorithm which detects pedestrians in video signs using different classifiers is presented. A special classifier based on the majority vote with the combination of support vector machines (SVM), K nearest neighbors (KNN) and AdaBoost algorithms has been used as a descriptor HOG to extract features of image during the this study. Motion estimation is implemented on the video image with three different classifiers to decide whether or not it is pedestrian. To demonstrate the effectiveness of the method, the system was trained with the INRIA data set and tested with Penn Fudan data set. It is clear that the proposed method decreases the false positive rate while increasing the true positive rate. Thus,it can be said that the majority voting method produces more positive results for the solution of the problem presented. The experimental results suggest that the method works with higher accuracy than pedestrian detection with a single classifier.