Discriminative Feature Learning For Person Re-identification

In recent years,with the growing security need of countries and individuals,the bigger the camera network is,the more difficult it is to find the key information from the massive video data efficiently.Therefore,person re-identification(Re-ID)has received a lot of attention from academia and industry,its main goal is to identify the captured pedestrians over different non-overlapping camera views.However,due to the factors such as lighting,camera angles and resolution,occlusion,background clutter,and diversity of pedestrian poses,there are still many challenges in Re-ID,especially on discriminative feature learning.For example,how to extract key information from video efficiently and tackle the common problem in Re-ID,person misalignment,are still activate tasks.To solve these issues mentioned above,the contributions of this paper includes:(1)As for video-based person Re-ID,limited by the research equipment,many algorithms are trapped in the calculation for the huge amount of data in the video.In order to improve the efficiency of video-based person Re-ID,this paper proposes to build a reinforcement learning based adaptive network to select discriminative frame sequences in the video,and feed the feature of the selected frame sequence to a pre-trained person Re-ID network,then the frame selection network will obtain feedback from the Re-ID network.The two networks are jointly trained and promote each other.The final experimental results prove that the proposed model can gain state-of-art performance with less computing resources on Re-ID.(2)As for image-based person Re-ID,person misalignment is a common problem in person Re-ID,because of the poorly detected bounding boxes or the variety of person poses.Obviously,misalignment will damage the accuracy and computational efficiency of person Re-ID.To solve this problem,this paper proposes a pose invariant deep metric learning method.First,we construct a pedestrian image Pose Box which shows a standard upright pedestrian pose.Then we input the original person image and Pose Box together to the network.The original image can retain the complete information of pedestrians to avoid the critical information loss,and Pose Box can utilize the pose to realize pedestrian alignment.At the same time,in order to avoid pose estimation error,the confidence score of pose estimation is the third input of the network as auxiliary information.In addition,to avoid a large intra-class similarity,we employ an improved triplet loss to improve the network performance.The proposed method has achieved pleasing performance improvements on different datasets.