Study On High-density Crowd Counting Algorithm Based On Convolutional Neural Network

With the rapid development of national economy and urbanization construction,the urban population is increasing rapidly.Crowd number is an important feature to describe group behavior.Crowd counting is very important for group monitoring and management,city planning,and traffic scheduling.High density in actual monitoring scenarios is faced with problems such as background interference and scale variation.The existing algorithms still face great challenges in the ability to resist noise interference,scale variation,and multi-task joint learning.To solve the above problems,the thesis studies the high-density crowd counting method based on convolutional neural network.The main research contents of the thesis are as follows:(1)To suppress background interference,a feature pyramid attention network(FPANet)is proposed.The FPANet firstly adopts lightweight dual attention module to reduce background interference.Then,the multiscale aggregation module adopts a learning-based cross-group strategy to aggregate and facilitate the fusion of feature maps along different channel dimensions.Extensive experimental results demonstrate that the FPANet achieves superior performances in terms of accuracy and efficiency under complex background interference.(2)To cope with scale variation,an attentive hierarchy Conv Net(AHNet)is proposed.The AHNet extracts hierarchy features by a designed discriminative feature extractor and mines the semantic features in a coarse-to-fine manner by a hierarchical fusion strategy.Meanwhile,a recalibrated attention module is built in various levels to suppress the influence of background interferences,and a feature enhancement module is built to recognize head regions at various scales.Experimental results on crowd datasets and cross-domain datasets illustrate that the AHNet effectively alleviate the effect of scale variation and achieve competitive performance in accuracy and generalization.(3)To synergize crowd counting and localization tasks and improve accuracy under largescale variations and background interference,a scale-context perceptive network(SCPNet)is proposed to jointly tackle the crowd counting and localization tasks in a unified framework.Specifically,a scale perceptive module with a local-global branch schema is designed to capture multiscale information.Meanwhile,a context perceptive module,by the channel-spatial selfattention mechanism,is derived to suppress the background interference.Furthermore,a hierarchical scale loss function that combines the Euclidean loss function and structural similarity loss function is designed to prompt the proposed model to fulfill the counting and localization simultaneously.Experimental experiments show that SCPNet can simultaneously improve the performance of high-density crowd counting and localization under scale variation and background interference.