Reseach On Human Identification Technology And Its Application In The Monitoring Of The Construction Site

Improving the safety management of construction sites has attracted more and more attentions along with the progress of China’s urbanization and increasing variety of infrastructure. Traditional camera surveillance cannot meet the requirement of safety management due to the complexity and personnel mobility of construction site. Therefore, it is important to address a smart camera surveillance system for construction site monitoring.In this thesis, we studied the human shape recognition technology in smart camera surveillance system, including three components: target recognition, feature extraction and classifier design. Firstly, we choose the suitable solution for human shape recognition by analyzing the difficulties in construction site scenario. Then, we studied the three major components: target recognition, feature extraction and classifier design separately. Finally we verified the effectiveness by simulation with construction site video. The outline of this paper is as follows:1) Target detection algorithm:By comparing and analyzing several popular target detection methods, we consider the background modeling method, which could separate the target against the background effectively, is suitable for construction site. We further present a target detection algorithm based on joint adaptive Gaussian mixture model and the space-time information of color histogram.2) Feature selection and optimizationWe summarized the characteristics of several existing features and build the eigenvectors using combination of features which are complement with each other. We reduced the dimension of vectors without loss of information by optimizing the eigenvectors using criterion function. Finally, we get a combination of feature which could describe humanoid target effectively without increasing the computation complexity significantly.3) Classifier designWe studied discriminant function theory. The linear classifier has simple form with low complexity, however, its classification accuracy is far worse than the non-linear classifier. We use piecewise linear classifier to obtain a trade-off between complexity and accuracy. We use both single-objective method as well as multi-objective method to guide the design of piecewise linear classifier. We address two piecewise lineae classifier which based on Max-min divisibility incremental design method and combination design of polyhedral separability, respectively.