Research On Dense Crowd Counting Based On Scale-aware And Attention Mechanism

The rapid advancement of urbanization has contributed to gradually colorful of people’s lifestyles.It is increasingly common for a large number of people to gather in public places in stadiums,concerts,shopping malls and other public places to participate in various activities.And it is particularly important to control the size of the crowd in these scenarios.In addition,the COVID-19 has swept the world in the recent years,and in order to prevent the spread of the epidemic,various public places are under strict control of the crowd gathering.Therefore,in order to ensure the safety of crowds in public places,it is of great significant to monitor and analyze the high-density scenes by means of crowd counting so as to conduct timely and effective evacuation of crowds.In recent years,with the emergence of new methods and models in crowd counting have been emerging,driving the steady development of this field.However,in practical application scenarios,the scale of crowds in scenes varies greatly due to different locations and angle changes of surveillance devices,which affects the effective extraction of crowd features at different scales at the same time.In addition,the image scenes captured by video surveillance are often complex and diverse,and there are objects similar to the head in the background information,which can easily be misidentified as crowds.And accurately distinguishing the foreground and background regions of crowds is also a problem that needs to be solved urgently.Therefore,this thesis conducts research on the above problem,and the main research work and innovations of this thesis are as follows:(1)Aiming at the problem of crowd head scale continuous variation in scenes,this thesis proposes an Attention-based Aware Pyramid Network.In this network,the aware pyramid module is designed to extract features on different scales by dividing the input features into blocks of different sizes and performing features fusion.Then the fused scale features are cascaded with the original input features to enhance the network’s ability to integrate features at different scales.In addition,the network designs the Spatial Attention mechanism(SA)and the Channel Attention(CA)mechanism.Specifically,SA selectively aggregates the features at each position by a weighted sum of the features at all positions,which is used to process feature information in the global context and captures density distribution in the feature maps.CA selectively enhances the important channels related to the crowd by dealing with the mapping relationship between any two channels while suppressing the channels with a large amount of interfering background information.Sufficient experiments on the public datasets of crowd counting demonstrate that the proposed method can not only significantly improve the accuracy of the crowd counting task,enhance the robustness of the network,but also better adapt to the continuous scale variations of the image.(2)Aiming at the problem of background clutter in complex scenes,this thesis proposes a Guided Axial-attention Multi-scales Aggregation Network.The network consists of two key modules: the Scale-aware Context Aggregation Module(SCAM)and the Guided Axial-attention Module(GAM).Specifically,SCAM progressively aggregates multi-scale contextual features by tightly connecting multiple dilated convolutions of varying receptive fields.Moreover,performing this operation on different convolutions layers not only achieves enhancement of multi-scale features on single-layer feature map but also capturing rich context and scale diversity.GAM is able to integrate local features with their corresponding global dependencies.And the attention loss is used to guide the axial attentional mechanism to ignore irrelevant information and emphasize to focus on the area of the image related to the crowd target.It not only contributes to alleviating the interference caused by cluttered backgrounds on crowd counting task,but also effectively reduces the computational complexity of the model.Experiments on challenging datasets demonstrate that the proposed method can improve the performance of counting while better adapting to a variety of complex scenarios.