Multi-task Study On Person Re-identification And Attribute Recognition Based On Local Features

With the development of society,people's living standards have been continuously improved,and more and more attention has been paid to safety and prevention issues.As an important measure of social security prevention,the surveillance camera network has also been vigorously built in recent years.This paper research two topics that apply to surveillance camera networks.Person re-identification refers to pedestrian matching across cameras.Pedestrian attribute recognition refers to the identification of the gender,clothing and other information of the pedestrian from pictures.Person attribute recognition by pictures taken by the surveillance cameras is very challenging.Different environments may result in low light intensity,low resolution,and partial occlusion of pedestrians in pictures.Due to the cross-camera pedestrian comparison,person re-identification task is often affected by factors such as the viewing angle changes and pedestrian pose changes that cause the body to be misaligned in the pictures.So how to extract robust pedestrian feature is the key to this research.field.This paper proposes a multitasking framework for person recognition and attribute recognition.The pipeline of the framework is:first reconstruct the human limbs in the picture through the pose estimation method.Based on this,the pedestrian image is divided into upper body and lower body pictures.The obtained two-part picture and the original picture are input to the trained local network and the global network,and the feature vectors are respectively extracted.At the same time,the attribute recognition branch in the global network will output the predicted value of each pedestrian attribute.The obtained featrue vectors are concatenated to obtain a joint feature vector.Finally,the results of the pedestrian recognition task are obtained by comparing the joint feature vectors.We conducted experiments on commonly used datasets.The method proposed in this paper can achieve good results in both the two tasks.Due to the integration of local features,the accuracy of the re-identification task can be improved.The multitasking framework of this paper saves a lot of time and storage space of network parameters for single task network training by multiplexing the feature extraction network.the two task branches do not interfere with each other and cause performance degradation.