Research On Population Density Estimation Algorithm In Embedded System

With continuous development of social economy,the level of town improves continuously.More and more urban population and crowds everywhere cause increasing abnormal group events,bringing people's lives and safety issues huge threat.In the places like railway and bus station where visitors flow rate are huge,potential safety hazard is especially much severer.As a result,estimating the crowd density with the help of video surveillance systems earns more and more attention.In recent years,there are a lot of achievements related to crowd density estimation algorithms,however,there are some inevitable shortcomings,such as huge calculated amount,poor performance of real-time and low precision.This paper summarizes the experience of some predecessors,comes up with some improvements in accuracy of algorithms and performance of real-time,and achieves crowd density estimation in embedded systems.The main work is as follows:Firstly,this paper simulates and makes comparison among four foreground extraction methods,which are background subtraction method,inter-frame difference method,three-frame difference method and background subtraction and modified three-frame difference method,analyzing their advantages and disadvantages at the same time.Then it extracts edge from extracted foreground image by Roberts operator,and eliminates the influences causing by mapping deformity using linear interpolation perspective adjustment method.The results show that the effect of extracting crowd foreground is impressive and the requirements of real-time are met.Secondly,the paper studies the population density classification algorithm based on texture feature analysis and uses GLCM to extract image texture features.What's more,it selects the appropriate orientation angles,grayscales and steps through simulation experiments.In order to further reduce the calculated amount,the paper proposes a reducing-dimension method based on principal component analysis,and obtains two principal components.After that,it extracts the most suitable kernel function and penalty factor for support vector machine by cross-validation method,which improves the accuracy of crowd density classification significantly.Experiment and simulation verifications are done at last.Finally,this paper accomplishes the transplant of GLCM algorithm to embedded systems through optimized design and implementation of data transmission,access bandwidth and pipelining operation.Validating experiments show that the transplant of algorithm satisfies the requirements of real-time.