Research On Some Key Technologies Of People Counting In Crowed Scenes

People counting is a fundamental and critical applied technology in the field of computer vision. Today, many public places are required to have a robust function of people counting in the video surveillance system. After years of researching and development, Recent robust algorithms of people counting are based on three low-level algorithms: pedestrian detection, pedestrian tracking and crowd density estimation. Due to the inherent complexity of human body and the practical scenes, there is no all-around algorithm of people counting which is robust and real-time. This makes people counting still an active and challenging region in the field of computer vision.Based on the current directions and the main challenges of people counting, the main contents of this thesis focus on some key technologies of people counting in crowded scenes. The major efforts of this thesis include the following items:1. Give a comprehensive introduction to the basic steps of traditional people counting including pedestrian extraction, classic pedestrian features, judgment of pedestrian and crowd density estimation. The introduction includes the basic principles, the applicable scenes and the algorithm performance.2. Extract pedestrian targets by background subtraction. We give a comprehensive introduction to the main and current algorithms of background modeling.In the research of Code Book method, we proposed an updating and elimination strategy of Code Word by setting weight. Experiments show that this weighted-Code Word can effectively improve the real-time performance of algorithm.3. We combined the HOG and the LBP into fusion feature, then through PCA method we send the feature vector into SVM to judge the result. Experiments show that the method can effectively improve the accuracy and the speed of pedestrian detection.4. We firstly predict the target through color histogram, at the same time we use pedestrian detection by HOG to modify the particle re-sampling. Through this method we comply the tracking task. Experiments show that pedestrian tracking based on particle filter and multi-feature has better accuracy and robustness comparing with the traditional methods.