Study On Crowd Counting Methods Of Complex Scenarios

The improvment of material life standard and the increasing number of culture activities bring people together more often.Due to the lack of effective early warning and monitoring system,large-scale crowding and stampede events occur from time to time,which poses a great threat to the safety of people.Traditional video monitoring systems rely on manpower,which is timeliness and effectiveness,while the development of computer vision counting and the upgrade of monitoring equipment have made it possible to automate crowd counting.The crowd counting system estimates the crowd density of various complex scenes in real time,providing early warning for possible casualty accidents and reducing the emergency rate of accidents.It is an important part of smart city and safe city.Traditional crowd counting methods can be divided into the following categories: dectection-based approaches and regression-based approaches.Both two kind of methods are highly depend on the choice of hand-made features and the classifier or regressor.Hand-made features are limited and hard to modified with respect to the specific problem,while the deep learning based methods provided a situation for that.The deep learning based methods realizes automaticly extracting and combining of effective features through the complex mapping function formed by the coupling between the layers,findind the related features which are more suitable for the crowd counting task,and achieving better performance.For the problem of high computational complexity and too many parameters in present multi-column density map estimatior(DME)network,this thesis analyzes the necessity of the structure through experiments.It turns out that the multi-column structure is not necessary.Based on that,we propose a single-column and deeper density map network with a Prior network branch to predict the couning number of crowd.By incorporating high-level prior information into DME branch,the joint regression module achieves a better performance.To avoid the false recognition and unsatisfactory performance of regression module in low population density,we introduce a detection module based on Faster R-CNN to realize the recognition of large objects to deal with the texture area where error recognition may occur in the regression module,which is complementary with the regression module.The output density maps from detection module and regression module are merged and sent to the fusion module to get the final output of our crowd counting system.Compared with other methods,experiment results on several typical data sets prove that our counting system proposed in this thesis has achieved improvement in both accuracy and robustness.