Research On Character Attribute Recognition Method For Surveillance Video

With the development of human science and social economy,people are increasingly exploring the information contained in various types of data.The surveillance scenario is an important research scene in the information age.The information mining of the surveillance scenarios also brings us a lot of convenience.Pedestrian attribute recognition is an important field of computer vision and pattern recognition.Recognizing the attributes of pedestrians is important for pedestrian tracking and pedestrian re-identification.In the traditional surveillance system,it is necessary to manually mark the attribute tags of the object or person that are interested in.So that,the subsequent tasks such as person retrieval can be continued.Obviously,this cannot meet the pedestrian attribute labeling requirements of the surveillance scenarios under big data.It is a wise choice to use computer vision to automatically and efficiently mark pedestrians in surveillance.In order to achieve this goal,this paper studies the pedestrian attribute recognition under the surveillance scenarios.This paper studies the pedestrian attribute recognition methods under the surveillance scenarios.The low-resolution problem and the occlusion problem of pedestrian images are also considered in this paper.The main contributions and innovation points have been shown in the following several aspects.Firstly,this paper studies the traditional attribute recognition method based on traditional hand-craft features.Based on this method,the input image are cropped into different strips to improve the results.Different attributes of pedestrian distribute in different parts of the pedestrian images.According to this,the attributes are divided into global attributes and local attributes.According to the different properties of attributes,the hand-craft features are extracted from different parts of the pedestrian images for attribute recognition.Excellent results have been achieved by this method.Secondly,this paper studies the pedestrian attribute recognition method based on ResNet network.This paper also respectively studies the attribute recognition methods that recognizes each attribute individually and recognizes all attributes jointly.In the method of recognizing each attribute individually,the recognition of each attribute is regarded as a binary classification problem and models are trained for each attribute.In the method of recognizing all attributes jointly,considering the internal relationship between different attributes,the recognition of attributes is regarded as a multi-label problem and one model is trained to recognizing all attributes at the same time.Obviously,the method of recognizing all attributes jointly achieves better result than the method of recognizing each attribute individually.Thirdly,this paper studies the effects of low-resolution and occlusion problems on attribute recognition in surveillance scenarios and the solutions for these problems.To solve these two problems,SRCNN is used to improve the resolution of lowresolution images,and the hash algorithm is used to recover the occlusion images in this paper.Experiments have proved the feasibility and effectiveness of these two methods.