Application Research Of Attention Convolutional Neural Network In Pedestrian Attribute Recognition

In the context of increasing emphasis on the construction of safe cities,smart cities and other projects,surveillance video has become one of the most widely used information carriers,playing an important role in information collection,mining and recording.Pedestrians as one of the important monitoring targets,if their attributes and other advanced semantic information can be effectively recognized in massive video images,which will greatly promote the development of security monitoring and other fields.However,the existing pedestrian attribute classification algorithm can usually only get better prediction results under certain specific conditions,for the appearance of pedestrians with large differences within the class,complex and variable surveillance video scenes,etc.,their robustness is constrained.Therefore,how to improve the generalization performance of the recognition algorithm and achieve accurate and efficient attribute resolution effect is still a very challenging topic.The main research contents of this article are as follows:1.Based on the existing target analysis methods,this paper analyzes the development status of pedestrian attribute recognition at home and abroad,and summarizes the current challenges and shortcomings of this task.2.Aiming at the problem that pedestrian attribute recognition is susceptible to changes in non-ideal conditions such as angle of view,scale,and illumination,and some fine-grained attribute recognition is difficult,a multi-level attention skip connection network is proposed.In the middle layer of the network,the sensitive attention module is used to filter and weight the original feature vectors in the channel and spatial dimensions,and the multi-level skip connection structure is designed to comprehensively consider the extracted salient features;at the top layer of the network,the multi-scale feature fusion method is improved to fuse local features and global features of joint optimization;at the network output layer,combined with the verification loss trend algorithm to adaptively update the loss layer,speeding up the convergence speed and accuracy of the model.3.The performance of the proposed attribute recognition algorithm is verified on the PET A and RAP datasets,and the effects of the designed multi-level attention skip connection module(MLASC),the improved multi-scale feature fusion method and the adaptive loss weighting strategy on the model prediction performance are further analyzedThe experimental results show that:(1)the multi-level attention skip connection network proposed in this paper has obvious advantages in recognition accuracy and model convergence speed,and has good generalization ability under non-ideal natural conditions;(2)the MLASC module and the multi-scale feature fusion method can effectively improve the network's robustness to fine-grained attributes and multi-scale pedestrian targets;(3)The adaptive loss weighting strategy can dynamically schedule different attribute tasks to accelerate model convergence...