Object counting for intelligent video surveillance

Abstract

The population of the world has been increasing and crowded scenes are more likely to occur, especially in some big cities. Accurate object counting is able to estimate the number of objects and avoid congested crowd causing accidents through reporting the real-time surveillance traffic of people condition to the security department. Moreover, it also can monitor the traffic flow. For places with large traffic flow, the Transportation Bureau can send more people to direct traffic, or to broaden long-term congested roads. Radio stations can give suggestions to drivers about which road is congested and which alternative way is recommended by receiving information from the monitors. Thus, crowd counting takes a more important role in every aspect. It is intuitive to think up that recognize people at first, which is a mature method. And then count the number of boxes enclosing people. It indeed effective in some scenes which are not congested. In a general crowded scene, people are likely to be hidden and obstructed by others, leading the computer cannot recognize people successfully or ignore the hidden persons. Convolutional neural network works well for regression or classification tasks, and it also proves its value in generating density maps. It builds an end-to-end regression method. This takes the entire image as input and directly generates a crowd count. In this dissertation, we introduce the basic theory about convolutional neural networks, analyze deep learning networks related with crowd counting and modify the model, algorithm and training details to improve the performance of the accuracy of crowd counting. Experimental results show that our method is effective to improve the accuracy of crowd counting.