Entrance People Counting System Design And Implementation

With the rapid development of China's economy, there are various large buildings, shopping and leisure facilities and the activities of people are becoming more and more frequent. How to monitor the activities of crowd in a large occasion to ensure the safety of the crowd, how to evaluate whether the facilities in buildings meet the customers' demands and how to distribute and schedule the public system resources, have become an important issue. At present, traditional and low-effective people counting system in entrance could not satisfy the demands of modern information technology, so the automatization and intelligence of people counting system is becoming more and more important.Through research and analysis, this dissertation combines with the advantages of moving target detecting and tracking technology as well as artificial neural network to carry out people identification and counting. This combined method can avoid the disadvantage of traditional moving target detection without the object recognition, which can't synchronously counting the moving people close to the body and may count interference objects mistakenly. At the same time, this method also can avoid the the disadvantage of artificial neural network without the target tracking, which may count the same target repeatedly.Firstly, the development of people counting system in domestic and foreign countries is introduced in this dissertation. It introduces the processing flow and the common key technologies of people counting system. In particular, it also analyzes and compares the common technologies of object detecting and tracing as well as the process of image recognition. On the basis of these, this dissertation researches the design of architecture and the process of implementation in actually, illuminates the function of each module and describes the running process of people counting system.Then, it researches the implementation of the image processing part in some depth. Especially, the dissertation researches the theory and implementation of the foreground detection module as well as the target tracing module. It also illuminates the related technology and the algorithm flow of foreground detection module. Besides, the target tracing module combines the Mean-Shift tracking algorithm and Kalman filter to improve the tracking precision.In the following, it researches the features of images which should be extracted. Based on these features, this dissertation also studies the theories and relative algorithm of the image recognition technology based on machine learning. Through the anysis of this project, the dissertation decides to choice the Feed-forward Neural Network using backpropagation algorithm. For the disadvantage of the BP algorithm, which may induce the problem of local minimum point and the slow speed of convergence, this project uses Levenberg-Marquardt (LM) algorithm based on optimization theory to improve the BP algorithm. Through the comparition of experiment, the BP network using LM training algorithm can converge fastly and also can decrease the training time. Then, it describes the four steps of building the neural network classifier.Finally, the conclusion and the prospect are also presented in this dissertation. Experiments demonstrate that the accuracy rate of system's recognition reaches above 98% and the system has good practical value.