Research On Pedestrian Target Alignment Method Based On Attention And Local Features

With the gradual development and optimization of deep learning technology in the application field of computer vision,pedestrian target alignment task no longer needs to manually label data information in various studies oriented to computer vision.At the same time,there are a large number of easily accessible pedestrian image data in daily life,and the application of these data has also attracted widespread attention in recent years.The pedestrian target alignment task aims to learn the features of a series of pedestrian target images,so that the model can automatically identify whether two pedestrian targets under different cameras are the same pedestrian.However,due to the low image resolution,pedestrian targets are blocked,different camera installation angles and scene changes,the pedestrian target alignment task is still very challenging.Although the research on pedestrian target alignment task at this stage has slowly changed from the method of extracting shallow features such as pedestrian clothing color and texture to the method of extracting deeper features,most of the current research results,especially the alignment research on the occlusion of pedestrian targets,usually still extract meaningless or even wrong features,This leads to problems such as the decline of alignment accuracy.Aiming at the task of pedestrian target alignment,this paper first studies and analyzes various mainstream pedestrian target alignment algorithms.Then,a pedestrian target alignment network using local features,Aligned RI network,is deeply discussed,and its alignment principle and existing problems are analyzed.Based on the Aligned RI network,this paper proposes an improved pedestrian target alignment algorithm based on joint relationship aware attention and adaptive relationship integration.The main contributions are as follows: 1)aiming at the problems of difficult target feature extraction under complex background and information overload caused by too much information in the image,this paper proposes a joint relationship perception attention module.Obtain the relationship between feature nodes in the two dimensions of space and channel,enhance the target area to be concerned,weaken the influence of interference factors such as background,and strengthen the ability to quickly obtain key features from a large amount of noisy information.2)To solve the problem of inaccurate local feature representation caused by pedestrian target occlusion,an adaptive relationship integration module is proposed in this paper.By calculating the difference between local features and global features,the weight matrix between local features is obtained to highlight more representative local features and enhance the subsequent local feature alignment ability,so as to achieve higher alignment accuracy.Finally,the improved algorithm and the current typical pedestrian target alignment algorithm are evaluated on Market1501 data set,Duke MTMC data set and CUHK03 data set.At the same time,the first matching rate and average accuracy are used as the evaluation indexes of the algorithm.In the Market1501 dataset,the first matching rate was 94.26%,and the average accuracy was 82.13%.On Duke MTMC dataset,the first matching rate was87.46%,and the average accuracy was 74.58%.On the CUHK03 dataset,the first matching rate was 71.62%,and the average accuracy was 73.97%.It can be seen from the experimental data that the pedestrian target alignment method based on joint relationship perception attention and adaptive relationship integration proposed in this paper shows good performance on multiple data sets.Compared with other alignment algorithms,the alignment algorithm in this paper can extract pedestrian target features with strong discrimination,obtain higher accuracy,greatly improve the alignment results and better performance.