Research And Application Of Crowd Counting Algorithm In Monitoring Scenes

Along with the continuous development of the social economy and computer technology, people pay more attention to the public safety problem. Nowadays, intelligent video surveillance system has been used in hospitals, shopping malls, schools, railway stations, bus stations, residential areas and other public places, which assist the security personnel timely to deal with emergency and guarantee the citizen’s personal and property safety. At the same time it can also help managers timely to tackle the problem of allocation of public facilities, which makes public resources reasonable to allocate and social life orderly and fast. Crowd counting has an important practical significance in the field of intelligent video monitoring, and is also a hotspot and difficulty in the fields of computer vision and pattern recognition. Estimating the number of people actually and real-timely in the monitoring scene can help some people advance warning and decision-making. In public security prevention and control and business information acquisition, Crowd counting system has very important research significance.The research status and characteristics of crowd counting algorithms are detailed analyzed at first. Then, these methods are classified and summarized. The crowd counting methods based on surveillance videos could be classified into two types, which are the method based on target detection and the method based on feature regression. These two methods have their advantages and disadvantages, and each of them has its applicable scene. The advantages of the method based on target detection are high accuracy of target position and high reliability of counting result while the disadvantage is unsatisfied detection effect in the case of complex environment or target occlusion. The advantages of the method based on feature regression is the robustness to the situation of dense crowd and the disadvantage is hard to find a common feature to fit the situation of complex background and camera perspective. Two different crowd counting methods of surveillance videos are proposed in terms of existing problems.In view of less occlusion and lower diversity of head in most monitoring scenarios, crowd counting algorithm based on the combination of Adaboost and kalman filter, which treats head as the crowd counting target, is proposed in this paper. First, head detection classifier is obtained through Adaboost algorithm and head is preliminary detected based on multi-scale sliding window and non-maximum suppression algorithm. Then, the images classified wrongly and ambiguously in the first step are selected to learn as the training samples deeply. Preminary detection result is filtered by the secondary classifier so that the errors can be removed effectively and the counting credibility is enhanced. Finally, the missed head is made up by kalman filter and data association algorithm. The result after the process above is treated as the counting result. This paper experiments with the record videos, and validates the robustness and effectiveness of the algorithm, and accuracy of crowd counting increased by more than 10% compared to preliminary detection results.The selection of features is the key of the crowd counting method, and the advantage of proposed method based on convolutional neural network is feature selection. The feature learned by deep learning is more in line with the process of biological recognition than other traditional features. We can get the crowd density map from regression of the centers of heads through convolutional neural network, then use ridge regression model to analyze the crowd density map to get the number of people in current frame. The method is proved to behave well in terms of counting accuracy and time-consuming computing procession, and the MAE has decreased by at least 6% compared to traditional methods.