Research On Anomalous Behavior Recognition Method And Technology For Public Safety

Anomalous behavior recognition for public safety is a crucial guarantee technology in the field of public safety risk prevention and emergency management.This dissertation aims to address the needs of public safety risk prevention and emergency response applications,by studying anomalous behavior recognition methods and techniques related to anomalous crowd aggregation analysis and abnormal vehicle re-identification for public safety risk prevention and control.The goal is to provide support technology for the development of intelligent,networked,and practical anomalous behavior recognition systems.This dissertation deeply analyzes the risks of anomalous crowd and vehicle aggregation in the context of public safety and investigates feature extraction and model building problems for anomalous crowd aggregation analysis and vehicle re-identification under constrained conditions,such as perspective effect,non-uniform crowd distribution,scale variation,occlusion,and illumination variation.The dissertation focuses on the design methods of abnormal behavior recognition deep network structure,density estimation,and loss functions,exploring ways to improve recognition performance under the influence of multiple factors.It innovates deep learning-based approaches for scene density perception,crowd counting,crowd localization,and vehicle re-identification.The main research work and innovative achievements of this dissertation are as follows:(1)In response to the issues of non-uniform crowd distribution and mutual occlusion in anomalous crowd gathering risk perception,an adaptive segmentation network scene density perception algorithm based on crowd counting is proposed.Firstly,a deeper network model is used to extract high-level,abstract feature information from images,which is combined with a spatial pyramid network and fused through a cascading manner to obtain more comprehensive context information.Secondly,a scene adaptive segmentation model is designed and trained to automatically extract discriminative features from scene images.Thirdly,a new regularization method is introduced to automatically find suitable nonoverlapping group assignments,improving the model’s detection ability for targets in different scenes.Finally,two crowd counting networks based on convolutional neural networks have been devised,employing deconvolution and dilated convolution to generate density maps for far and near regions,respectively,for crowd count regression,thus augmenting crowd counting precision.Experimental results show that the proposed adaptive segmentation network scene density perception algorithm based on density levels has high counting accuracy and is suitable for complex scene density perception tasks.(2)In response to issues of large head size changes in crowd counting tasks in highly crowded scenes of anomalous crowd gatherings,a crowd counting algorithm based on multiscale feature adaptive fusion networks is proposed.Firstly,a truncated deep network model is adopted to design different level feature extraction methods,retaining the three largest pooling layers to enrich features at different scales.Secondly,a hybrid attention mechanism of light weight is then presented to diminish feature information loss due to channel rivalry,accentuating efficient data,suppressing superfluous data,and augmenting the expressiveness of feature maps.Finally,a dilated convolution module is designed,combining traditional convolution and dilated convolution to accelerate network convergence,thus generating highquality estimated density maps.Experimental results show that the proposed crowd counting algorithm based on multi-scale feature adaptive fusion networks further improves the accuracy and robustness of crowd counting,and better adapts to complex situations such as head size changes in highly crowded scenes.(3)In response to issues of anomalous crowd analysis tasks,a single model cannot uniformly model different types of relationships.A weakly supervised crowd analysis method based on local receptive fields is proposed.Firstly,the Transformer is introduced to the crowd analysis task,designing a weakly supervised crowd analysis framework based on local receptive fields,achieving crowd counting and crowd positioning functions.Secondly,a head detail enhancement method based on global max pooling is adopted to extract multi-scale features from images and retain more head detail information.Thirdly,a new binarization network is explored,equipped with an adaptive threshold learner,reducing extra supervision labels and obtaining accurate confidence binarization maps.Finally,a connected component detector is proposed for head position detection and initial box,improving crowd counting and positioning accuracy.Experimental results show that the proposed weakly supervised crowd analysis method based on local receptive fields achieves unified modeling of crowd counting and positioning and further enhances the algorithm’s cross-domain feature learning capability.(4)In response to the unique challenges and rapid re-identification requirements of abnormal vehicle risk perception caused by vehicle exterior damage and inability to identify vehicle identity through license plates,a lightweight abnormal vehicle re-identification method based on a shuffled vision transformer network is proposed.First,this dissertation employs a deep neural network model,integrating ShuffleV2 as the fundamental convolutional module,reducing network parameter size,lowering the computation required for feature extraction,and increasing re-identification speed.Second,this dissertation designs a ShuffleViT module that combines self-attention mechanisms,extracting semantic information from vehicle images while preserving spatial relationships between image features,enhancing the model’s resolution for global context information and local detail information.Third,this dissertation introduces a supervised contrastive loss function,which,combined with a cross-entropy loss function,is used for joint training.This not only increases the feature distance between different vehicle images but also strengthens the constraint on the feature distance within the same vehicle images,improving the model’s discriminative ability.Experimental results show that the proposed lightweight abnormal vehicle reidentification method based on the shuffled visual transformer network can effectively focus on subtle differences between vehicle images,satisfying the requirements for rapid reidentification of abnormal vehicles.