Fine-grained Pedestrian Attribute Recognition In Surveillance Scenario

Computer vision is one of the powerful driving forces of artificial intelligence,its research goal is to enable computers to capture and understand visual data by imitating human beings.In recent years,with the development of security industry,monitoring equipments have grown rapidly.It has become a hot research field to use computer technology to realize the intelligence of monitoring system.Pedestrians are one of the most important monitoring targets.People are eager to acquire the visual attributes(somatotypes,clothing styles,accessories and so on)of pedestrians in surveillance video.The recognition of attributes has become a research hotspot in computer vision.As the appearance features of pedestrians,efficient recognition of pedestrian attributes has been a basic task of pedestrian visual analysis,which is widely used in intelligent monitoring,human-computer interaction and image retrieval.It has enormous academic research and commercial application value.In the past ten years,attribute recognition technology continues to develop,and many excellent pedestrian attribute recognition algorithms have been proposed.However,most algorithms can only get better recognition under certain conditions.For variety monitoring scenarios,complex pedestrian appearance and other issues,its robustness is difficult to guarantee.Therefore,improving the performance of attribute recognition algorithms is still a challenging task.This thesis focuses on pedestrian attributes recognition in surveillance scenarios.The main work is as follows.1.A pedestrian attribute recognition method based on multi-stage learning and multiple loss function is proposed.At present,many pedestrian attribute recognition algorithms only consider the positive correlation among attributes for attribute reasoning,but ignore the negative correlation.It is still an open problem to explore the negative correlation among attributes explicitly.In multi-attribute joint training,the performance of some attributes will be affected by other attributes.The correlations can be divided into the following two categories:(1)mutual promotion,(2)mutual suppression.The former has a positive impact,which can improve the performance of some attributes of the model;The latter has a negative impact,which will make the performance of some attributes be suppressed from each other.To model such correlations into the learning pipeline,attributes need to be grouped according to the learning situation.The positive and negative correlation attributes should be grouped into the same and different group respectively.In this thesis,a deep learning based multi-stage pedestrian attribute algorithm is proposed to explore the positive and negative correlation simultaneously and jointly recognize all attributes.Moreover,a new loss function is proposed to further improve the prediction performance.The performance superiority of the model is demonstrated by comparing with twelve related methods.2.A multi-task pedestrian attribute prediction method based on multi-scale feature fusion and Bi-directional recurrent neural network(Bi-RNN)is proposed.There are two key points in the pedestrian attribute recognition:(1)The size of image region corresponding to different attributes is different.Therefore,multi-scale features need to be extracted;(2)mining the relationship between attributes to support the inference learning of attributes.In order to meet the first key point,the wider network,inception structures are used to extract features which have different receptive fields to enhance the semantic expressiveness.To further enhance the robustness of features,the features from different layers are fused to acquire multi-scale features.In order to achieve the second key point,The proposed model contains two sub-modules,i.e.the multi-branch network module and correlation exploration module.The former is utilized to capture the unique features of each attributes on the basis of the multi-task learning.Given the unique features,the later uses Bi-RNN to mine the relations between different attributes.Hence,the model can learn the features that conform to the relations among attributes.Two modules complement each other.Finally,the proposed model can improve the final performance significantly due to the attribute-specific feature learning and correlation mining.The comparative experiments show the performance superiority of this method.