Target Similarity Measurement Algorithm Based On Wasserstein Distance

In recent years,with the continuous development of society and the continuous advancement of information technology,human food,clothing,housing and transportation have become more intelligent,and camera surveillance systems have become increasingly common in the living environment.Pedestrian re-identification and multi-target tracking as key technologies are of great significance for the intelligentization and application of video surveillance systems.This paper mainly studies the application of Wasserstein metrics in the similarity measurement problem in pedestrian recognition and the data association problem in multi-target tracking.This paper proposes a pedestrian recognition algorithm based on Wasserstein metric and a data association algorithm based on Wasserstein metric in pedestrian tracking.The specific content includes the following two aspects:Pedestrian re-recognition algorithm based on Wasserstein metrics: Firstly,the Wasserstein distance solution step is deduced in detail.By means of Farkas lemma and duality theorem,the problem of finding the minimum Wasserstein distance is transformed into the maximum value of the dual form under constraint conditions.Then,combining Wasserstein distance with deep network,considering the difficulty and convergence of network training,a gradient penalty term is added to the output of the network to limit the output of the network to a certain range.Finally,experiments were carried out on the pedestrian recognition dataset Market1501.The experimental results show that the proposed algorithm has improved the cumulative matching characteristics(Top-1)and average accuracy(m AP)compared with other related algorithms.The data association algorithm in pedestrian tracking based on Wasserstein metric: First,the data set is made based on the target detection result on the MOT16 training set.The data set has a total of 110,000 data,each of which contains two 128-dimensional vectors,which respectively describe the appearance characteristics of the front and rear frame pedestrians.Then,based on the Wasserstein distance,a Contrastive loss is defined and trained on the data set that I made.The algorithm cascades the appearance and motion characteristics of pedestrians.The Wasserstein distance reflects the appearance matching of pedestrians.The Mahalanobis distance reflects the motion matching of pedestrians.For pedestrians who satisfy the appearance matching degree,the motion matching degree is further filtered.For pedestrians who simultaneously satisfy the appearance matching degree and the motion matching degree,the Hungarian algorithm is used to achieve the best correlation of pedestrians.Finally,experiments were carried out on the MOT16 testset.The experimental results show that the proposed algorithm effectively reduces the number of missing pedestrians in the tracking.