| Throughout the investigation practice,the strong ability of video surveillance to restore crime scenes makes it play an important role in many situations.As an important part of video inspection and identification,human image identification have also been extended from simple face comparison to various features related to human image,among which the spatiotemporal association feature of the human body,as a high-order feature that contains both the characterization feature and the temporal and spatial connection,can make up for the impact of the lack of face information on the weakening of the probative power of human image identification,and expand the identification angle.However,at present,relevant research is limited by the small scale of available data,and often can only summarize or put forward certain hypotheses in a small range,so it is necessary to use automated methods to statistically the required features of large-scale and multi-camera data to support the development of related research.In view of the above problems,the following research work is carried out:Firstly,a pedestrian attribute recognition algorithm combining semantic and image information is proposed to address the problem of insufficient extraction of natural semantic associations between pedestrian attributes and image information related to different attributes.Utilize self-attention mechanism to mine the relationships between pedestrian attributes,and utilize cross self-attention mechanism to combine image information with pedestrian attribute information to improve the utilization efficiency of pedestrian attribute information.Then,the design module integrates high-order features with low-order features,and adds local information through convolution calculation to enhance the generalization ability of the algorithm.The experimental results show that the algorithm in this chapter achieved 84.04%,79.71%,and 88.03% results for the mean accuracy,accuracy,and F1 indicators on the PA-100 K dataset,and 89.04%,82.39%,and 89.06% results for the mean accuracy,accuracy,and F1 indicators on the PETA dataset.Secondly,to address the problem of poor recognition performance caused by a small number of positive samples for some pedestrian attributes,a pedestrian attribute recognition algorithm based on fusion representation learning is proposed.Comparative learning is used to learn the representation features of images,improving the algorithm’s ability to recognize underlying features,and thereby enhancing the ability to recognize pedestrian attributes with fewer positive samples.At the same time,a weighted fusion strategy is designed to adaptively weight the loss function of supervised learning and representational learning,so that the algorithm can have different learning tendencies in different training stages,and improve training efficiency and recognition results.The experimental results show that the algorithm in this chapter improves the mean accuracy metric by 3.5% compared to the baseline on the PETA dataset and 3.18% compared to the baseline on the PA-100 K dataset.Thirdly,to address the degradation of model recognition ability between different datasets and meet the subsequent research needs of human spatiotemporal association,a cross domain pedestrian attribute recognition algorithm based on domain adaptation is proposed.Firstly,the problem definition is given to the requirements in real scenarios.Then,a domain alignment module is designed using cross self-attention to achieve domain alignment between different domains,and a metric covariance measure is introduced to transform the feature maps.The similarity between the two transformed feature maps is calculated to optimize the domain alignment effect and measure it.The experimental results show that both the unsupervised domain adaptive experimental setup and the less labeled experimental setup defined in this chapter have the best performance.This article proposes two methods to improve the recognition effect of pedestrian attribute recognition under monitoring,and proposes a feasible solution for cross camera recognition in subsequent research on human spatiotemporal correlation features. |
