Research On Crowd Counting Method Based On The Density Map Estimation For Complex Scene Images

Crowd gatherings become more frequent in the public places including the transportation,large event site and large shopping mall,and the consequent security risks are becoming more serious.It is essential to monitor the crowd count of the scene for the practical application,such as the video surveillance,traffic management,emergency system,and so on.Therefore,how to study the robust crowd counting method to make it more suitable for the unconstrained complex scene image of the real world is still important and difficult point of the complex scene image crowd analysis,scene understanding and video surveillance.To this end,this dissertation aims to study the single-image crowd counting method in the complex scenes based on the density map estimation,and the main research contents are as follows:Firstly,aiming at the issue of the uneven crowd distribution in complex scene images,the single-image crowd counting method based on adaptive feature aggregation is developed.Based on the multi-task learning idea and the hierarchical working principles of the convolutional neural network as well as the attention mechanism,the global feature information is modeled;according to the complementary feature fusion and the attention mechanism,the local feature information is extracted to aggregate the feature representation of the crowd region;the adaptive feature information is formed with the global and local feature information to enhance the feature representation of the counting model so as to reduce the negative effect of uneven crowd distribution on the counting accuracy.Secondly,aiming at the issue of the crowd scale variance in complex scene images,the single-image crowd counting method fusing the scale-aware information is proposed.Based on analyzing the characteristics of the crowd scale variance across the scene,the mapping from the crowd density degree to the crowd scale is established,and the crowd density degree prediction result is utilized to guide the crowd counting model to extract the scale-aware information for the current scene,to reduce the influence of the crowd scale variance across the scene on the counting performance;based on analyzing the characteristics of the continuous crowd scale variance within the scene,the an abundant scale-aware information extrication method based on the flexible network depth is investigated,which establishes the multi-branch path dependency for the multi-column convolutional neural network to realize the flexible network depth so as to extract the abundant scale-aware information from the crowded scene,decreasing the interference of the crowd scale variance within the scene on the counting accuracy.Then,aiming at the issue of the complicated background information in complex scene images,the single-image crowd counting method combining the foreground prior information is proposed.Taking into account the sensitive characteristics of the wavelet transform to the transient signal,the foreground contour prior information are generated and optimized using the wavelet transform and the convolutional neural network;based on the analysis that the thermal images could distinguish the foreground crowd and the background information,a non-linear mapping based on the modal transform is established to extract the internal features as the foreground region prior information;the crowd counting method is involved with the foreground contour prior information as well as the foreground region prior information,to alleviate the inhibition of complicated background information on the counting performance.Finally,aiming at the issue that the common Euclidean distance supervision item is unable to represent the correlation between the pixels,the single-image crowd counting method guided by the multiple supervision items is studied.According to the characteristic that the dependent relationship exists among the pixels of the crowd density,the supervision items guided by the full-reference and no-reference image quality assessment are studied to mine the structural relation between the pixels of the density map;in view of the phenomenon that the crowd counting result in the region is determined by summing the density map,the pyramid region-wise guided supervision item is investigated to mine the association relationship of the pixels of the density map;combining the Euclidean distance and the image quality assessment as well as the pyramid region guided supervision items,multiple supervision items with different assessment dimensions are formed,to enhance the accuracy of the single-image crowd counting method based on the density map estimation.