6533b856fe1ef96bd12b373a

RESEARCH PRODUCT

Real-time implementation of counting people in a crowd on the embedded reconfigurable architecture on the unmanned aerial vehicle

subject

description

The crowd counting task is an important research problem. Now more and more people are concerned about safety issues. Considering the scenario of a crowded scene: a population density system analyzes the crowds and triggers a warning to divert the crowds when their population density exceeds a normal range. With such a system, the incident of the Shanghai New Year's stampede will not happen again. The most difficult problem of population counting at present: On the one hand, in the densely populated area, how to make the model distinguish human head features more finely, such as head overlap. The second aspect is to find a small-scale local head feature in an image with a wide range of population density. The most critical aspect, in some public places, is that we can not install an intelligent video surveillance system. So how do we estimate the high-density crowd area to avoid crowd trampling accidents? Facing these challenges, we propose implementation of real time reconfigurable embedded architecture for people counting in a crowd area. First, our work integrates the features of HOG and LBP, which not only combines the effective identification information of multiple features, but also eliminates most of the redundant information, thereby realizing effective compression of information, saving information storage space. Then, in terms of crowd counting, we use multiple sources of information, namely HOG, LBP and CANNY based filtering. These sources provide separate estimates of the number of counts and other statistical measures, through the support vector Machine SVM, classification. At the same time, in order to effectively solve the problem of extracting scale-related features in crowd counting. We propose a new framework M-MCNN based on MCNN for crowd counting on any single image. M-MCNN not only contains the original three columns of convolutional neural networks with different filter sizes, but replaces the fully connected layers with a convolutional layer of 1*1 filters, so the input image of the model can be of any size. Moreover, in a single individual sample, we greatly improve the learning of sample features by extracting the texture features of a single human head , and better use it for datasets. Finally, we implement our new framework M-MCNN through FPGA, and transplant it on the drone to estimate and predict the high-density crowd area in real time. Our model achieved good results in crowd counting.