Open AccessResearch on Person Reidentification Method Fusing Direction Information in Multi-camera Pedestrian Tracking Problem
Author(s) -
Junliang Liu,
Xun Shi
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1871/1/012068
Subject(s) - pedestrian , computer science , normalization (sociology) , artificial intelligence , computer vision , feature (linguistics) , feature extraction , pedestrian detection , tracking (education) , pattern recognition (psychology) , engineering , transport engineering , psychology , pedagogy , linguistics , philosophy , sociology , anthropology
In multi-camera pedestrian tracking, pedestrian apparent features are usually used to solve the problem of cross-camera reidentification. In order to make the pedestrian apparent features robust to the directional changes of pedestrians in different cameras. In this paper, a pedestrian feature extraction method combining directional features is proposed by using deep network Resnet 50. Pedestrian feature extraction network is trained by adding pedestrian direction information. in order to improve the performance of person reidentification model, BNblock structure is added in the design, a batch normalization (BN) layer is added after the feature to get the normalized feature. It can make the triplet loss converge with the convergence of the ID loss, thereby improving the performance of the model. The proposed method was verified on DukeMTMC and Market-1501 data sets, and the results show that the pedestrian appearance features combined with direction information proposed by the pedestrian reidentification algorithm can significantly improve the reidentification accuracy. By adding BNblock into the network structure, the accuracy can be further improved.