| The goal of person re-identification(Re-ID)is to determine whether the target person appeared in the same camera at a different time or in another camera at a different time.The target person can be represented by an image,video sequence or textual description.In recent years,deep learning has succeeded in computer vision.As a result,deep-learning-based person Re-ID has also been widely used in smart security and smart city.Person Re-ID technology has different core problems in two types of objects,namely,image and video.The core problem of image-based person Re-ID is how to extract robust pedestrian features,while the core problem of video-based person Re-ID is how to extract the feature representation of the spatial-temporal information of a video sequence.In addition,different core problems must be addressed under fully supervised and unsupervised approaches.The core problem of unsupervised person Re-ID is generating high-quality pseudo-labels more efficiently with no labels.Based on deep learning technology,this paper proposes methods of person Re-ID and unsupervised person Re-ID for the two different object types(image and video).The main innovations are as follows:(1)For image-based person Re-ID,it is challenging to select effective features of the person Re-ID framework owing to the influences of camera angle,pedestrian posture,occlusion and background.Aiming at the problem of single effective feature selection in deep-learning-based person Re-ID,this work proposes an adaptive spatial scale(ASS)person Re-ID framework.ASS strategy first extracts the local and global features of Res Net and then aggregates features by adaptive scale to improve feature learning ability.The framework embeds the ASS strategy into the last two layers to achieve end-to-end training and high efficiency.Experiment results demonstrate that the method can improve the discriminative ability of the feature representation and resolve the inconsistency between ID loss and triplet loss.(2)A person Re-ID network with enhanced attribute awareness is proposed in consideration of the insufficient scale of the image-based person Re-ID dataset,low model generalisation ability of the deep learning network and the over-fitting tendency.The network includes global,local,and attribute awareness modules.The network exploits a non-local attention module to enhance the spatial semantic information of Res Net.A local branch module is designed to learn fine-grained local information effectively.A standardised processing module is also designed for each branch module to avoid excessive attribute learning.Experiment results demonstrate that the method effectively enhances spatial semantic information and achieves good recognition rate and accuracy.(3)Given that video-based person Re-ID problem is rich in spatial and temporal information,extracting effective features may be difficult.Thus,a video-based person Re-ID model based on multi-level fusion temporal-spatial co-attention(MLTS)is proposed in this work with the aim of solving the problem of easily lost spatial-temporal cue information in the image frames of pedestrian video sequences.The network includes attention,global and local modules.A local branch network is designed to avoid the misalignment of image frame posture in video sequences.A backbone network based on knowledge evolution is introduced to improve the model learning ability of Res Net.Furthermore,with MLTS,the network learns spatial-temporal cue.Experiment results demonstrate that these modules complement each other and work well together.The method can solve the problem of frame misalignment and reduce the interference of occlusion and chaotic background.(4)An unsupervised person Re-ID model with enhanced camera awareness is proposed with the aims of mitigating the difficulty of labelling large-scale pedestrian data and solving the problem of clustering method introducing noise labels.Pseudo labels are used for sampling to improve diversity and prevent over-fitting.The camera awareness module learns invariant person features and achieves significant performance by combining with a feature extraction strategy.A unified update strategy,which combines contrast and metric learning,is likewise proposed to learn a cluster-level memory dictionary better and dynamically update cluster features.Experiment results demonstrate that the algorithm can effectively reduce the influence of pseudo label noise and quickly improve existing methods’ performance.In summary,this paper conducts an in-depth analysis of several key technologies for deep-learning-based person Re-ID.It optimises CNN feature aggregation,adaptive attention to fine-grained local attributes,video sequence spatial-temporal alignment learning and camera awareness enhancement learning.The study of these methods is helpful in the development of person Re-ID technology and has specific theoretical significance and practical value. |
