Research On Multi-scale Crowd Counting Method For Anti-noise Interference

With the acceleration of the construction of smart cities in our country,the intelligent application of information technology has played an increasingly important role in social development.As a key component of the security monitoring system,crowd counting has drawn much attention from researchers in recent years.Hence,a technical issue,faced by both industry and academia,is how to improve the counting accuracy in surveillance video images.Crowd counting tasks face great challenges in real-world applications.For instance,there are many external factors such as changes in image perspective of crowd gathering places,complex background noise and diverse crowd distribution etc.,which seriously affect the accuracy of crowd statistics.In addition,due to the quality limitation of label samples,crowd counting algorithms based on convolutional neural networks are also difficult to achieve good performance.This paper focuses on the problems of scale changes,image background noise interference and label noise interference in crowd counting tasks.Specifically,our approach designs two crowd counting network and two training loss functions based on the self-adaptability of supervisory signals,which could improve the accuracy of counting performance in complex scenes with high perspective.Our work mainly includes the following research:(1)Aiming at the problem of scale diversion,a strong information perception model is proposed: Hierarchical Information Sharing Network(HISNet).Based on the feature fusion strategy,the baseline model is optimized by splicing multi-stage features with information sensitivity at different scales,which reduces the loss of model transmission information while improving the model’s multi-scale detail perception ability.Furthermore,the accuracy of the Gaussian probability regression of the crowd head is improved by enhancing capabilities of feature extraction;(2)Aiming at the problem of label noise interference,under conditions of noisy label,this paper studies to train crowd counting models through weak supervision signals and proposes a weakly supervised learning strategy.By aggregating the total number of people in the sub-scale region to coarsen the granularity of labels,the model can avoid fitting the pixel-level noisy label.However,the simple weak supervision source does not include enough information to train a model with good performance.Hence,based on the above strategy,two loss functions are designed: dynamic pyramid loss function(DP Loss)and regional pyramid segmentation loss function(RPS Loss).Among them,DP Loss updates the supervision label constraint granularity through the model training epochs,which could guarantee the fitting ability of model data samples while enhancing the generalization performance of the model;According to the image population density,RPS Loss firstly divides the multi-level density regions with different label noise rates,and then adopts the corresponding granularity supervision constraints for each density level region.The above setting can ensure that the model can be adaptively adjusted to the accuracy of multi-scale head feature fitting,and also enhances the generalization ability of the model.(3)Aiming at the problem of background noise interference,based on the above HISNet,a counting model is further proposed: the path correction network based on residual attention(RA-PCNet).The residual attention mechanism,contained in our methods,is utilized to refine the population distribution information in sharing features at all levels through the vertical path connection of high semantic features,which could reduce the interference of the complex texture background on the population count density regression,and further improve the accuracy of the crowd counting model.Experiments have been carried out for the above research work to verify the effectiveness of each method against background noise interference and anti-tag noise interference.Experimental results show that the accuracy of multi-scale crowd counting method for anti-noise interference proposed in this paper has achieved well-performance on two datasets.