Research On Crowd Counting Algorithm Based On Lightweight Network

With the acceleration of contemporary urbanization,intelligent monitoring has become an important means to maintain order in public places and ensure the safety of people’s property.As an important part of intelligent monitoring,crowd counting plays an important role in controlling the density of people in public places and strengthening epidemic prevention and control.In recent years,with the introduction and development of deep learning,crowd counting algorithms based on convolution neural networks have made continuous breakthroughs in counting performance and achieved outstanding results.However,in practical application scenarios,existing methods still face severe challenges,such as drastic changes in crowd scale,complex background noise,large number of model parameters,and long inference time.In order to improve the accuracy and practicability of the crowd counting algorithm,this paper studies the algorithm structure and proposes targeted solutions.The main innovations are summarized as follows:(1)Lightweight Multi-Scale Adaptive Convolutional Neural Network for Dense Crowd Counting(LigMSANet).Aiming at the problems of dramatic changes in crowd scale and poor real-time performance of current counting algorithms,a multi-scale adaptive lightweight crowd counting algorithm was proposed.First,this algorithm uses the lightweight convolutional neural network MobileNetV2 framework as the baseline to extract low-level crowd features.Through its test analysis,the first five bottleneck blocks are used,and some framework settings are changed,which makes it more suitable for the field of crowd counting;secondly,a multi-scale adaptive module MSAM(Multi-Scale Adaptive Module)is designed by adopting multi-scale fusion and scale adaptation strategies to break the feature extraction.It generates scale attention information according to the input image,and guides how to focus on the distribution of representations using different sizes of convolutions in the channel dimension to automatically adjust the size of the receptive field to improve the extraction of multi-scale features.Experiments on three main crowd counting datasets,Shanghai Tech,UCF＿CC＿50 and UCSD,show that compared with the state-of-the-art lightweight crowd counting algorithm,this method reduces MAE by 13,RMSE by 25,and achieves the fastest inference speed 42 ms.(2)A Real-time Crowd Counting Network Integrating Multi-scale Perception and Background Suppression(RMPBSNet).Scene complexity is another major problem in the field of crowd counting.In order to effectively reduce the impact of complex backgrounds on counting accuracy,without significantly improving the parameters and inference time of the model,this algorithm proposes a real-time scale based on LigMSANet.A crowd-counting algorithm for fusion of perception and background suppression.First,the MSAM module in LigMSANet is improved,and parallel global max pooling and global average pooling are used to extract features when generating scale attention information.Second,a lightweight background suppression module LBSM(Lightweight Background Suppression)is designed.Suppression Module),which utilizes rich low-level features to generate spatial attention maps to reduce the impact of background noise.The experimental results show that,compared with the existing crowd counting algorithms,RMPBSNet can extract robust multi-scale features and effectively suppress background noise,and has good competitiveness in both speed and accuracy.The paper involves 28 figures,28 tables,and 80 references.