Research On Crowd Counting Problem In Complex Scenes Based On Deep Learning

Due to the ongoing advancement of urbanization and the accelerating population growth,the problem of crowd counting has become a crucial issue in contemporary society,particularly in the areas of urban traffic,safety monitoring,and crowd control.In recent years,deep learning has demonstrated great potential in numerous computer vision tasks.However,in complex scenes with dense crowds,severe occlusion,individual variations,and large pose changes,existing crowd counting methods suffer from limited counting accuracy.Therefore,this thesis aims to investigate a deep learning-based approach to crowd counting in complex scenes by utilizing convolutional neural networks to learn advanced features and improve counting accuracy.Specifically,the main research work of this thesis is as follows.(1)This thesis proposes a multi-scale Faster R-CNN based crowd counting method using detection-based deep learning approach,addressing the limitation of the classic Faster R-CNN algorithm that only performs predictions on a single scale feature map.Specifically,CSPResNet+FPN is employed as the feature extraction network to generate feature maps of different scales.In the RPN stage,anchors of corresponding scales and aspect ratios are designed based on the different feature maps to adapt to multi-scale targets in complex crowd scenes.Additionally,RoI Align is adopted in the RoI Head stage instead of RoI Pooling to reduce information loss and improve detection accuracy.(2)In response to the problem of scale variation in complex crowd scenes,this thesis proposes a crowd counting method called CL-DCNN based on cross-layer contextual feature connections.Specifically,the method utilizes dilated convolutions to construct a dilated contextual module(DCM)with two branches.One branch effectively acts as a bottleneck to facilitate effective cross-layer connections of contextual features,while the other branch captures multiscale information using dilated convolutions.By stacking multiple DCMs at the back end of the CNN backbone network,CL-DCNN enhances the network’s ability to strengthen the flow of information between different feature layers,recognize local details and global context,and thus improve counting accuracy.(3)In response to the problem of severe background interference in complex crowd scenes,this thesis proposes a crowd counting method called DA-DCNN,based on dual-attention feature fusion,which improves upon method(2).Specifically,this method employs spatial and channel attention to weight different regions and channels of the feature map generated by DCM.Spatial attention enhances the network’s focus on the crowd region while suppressing interference from complex background noise,while channel attention strengthens the expression of useful channels by exploiting the dependency relationships between different channel features and suppressing the influence of useless channels.DA-DCNN uses the crowd attention map to guide the network in generating a higher quality density map,thus improving counting accuracy.In conclusion,this thesis proposes three different deep learning-based crowd counting methods to address the problem of crowd counting in complex scenes.The effectiveness of the proposed methods is validated through comparisons with other methods on multiple datasets.The research findings of this thesis are expected to provide strong technical support and references for solving crowd counting problems in real-world scenarios.