Method for generating and expanding pedestrian re-identification data based on generative network

Abstract

The invention provides a method for generating and expanding the pedestrian re-identification data based on a generative network. The method includes the steps of generating a new pedestrian video frame sample by utilizing a video prediction network; generating the end-to-end pedestrian background transformation data by means of a deep generative adversarial network; expanding the breadth and therichness of the pedestrian data set by using different data generation methods; and sending the expanded data set to the feature extraction network, extracting the features and evaluating the performance through the Euclidean distance. According to the method, intra-class and inter-class data expansion of pedestrians is taken into consideration as well, more abundant samples can be generated by combining different generative networks, the expanded data set has good diversity and robustness, the problem of performance loss caused by insufficient number of samples and background interference canbe well solved, and the method has general applicability, and the expanded data set can achieve better performance and efficiency in the next step of pedestrian recognition.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and specifically relates to a method for generating and expanding pedestrian re-identification data based on a generative network, in particular to a method for generating and expanding data suitable for improving recognition performance in pedestrian re-identification. Background technique [0002] Pedestrian re-identification is a key task in intelligent video surveillance, and it has been a research hotspot in the field of computer vision in recent years. It is suitable for technical fields such as security and public places. Pedestrian re-identification can be defined as: In a non-overlapping video surveillance network, for a given pedestrian in a camera, determine whether it appears in other cameras. It is an automatic target recognition technology that can quickly locate human targets of interest in the surveillance network, and is an important step in applications such as intellige...

AI Technical Summary

Problems solved by technology

Since the existing pedestrian re-identification datasets are generally too small to meet the training requirements of deep networks, it is often more likely to lead to over-fitting and performance loss

The existing methods are divided into three types. One is to collect more labeled data to expand the data set (see J.Ponce, T.L.Berg, M.Everingham, D.A.Forsyth, M.Hebert, S.Lazebnik, M. Marszalek, C.Schmid, B.C.Russell, and A.Torralba.2006.Dataset Issues in ObjectRecognition.Springer Berlin Heidelberg.29-48pages), but this method is too expensive, and the reliability of the labeling work is not high enough; the second is Add unlabeled data and perform unsupervised learning to improve performance (see Peixi Peng, Tao Xiang, Yaowei Wang, Massimiliano Pontil, Shaogang Gong, TiejunHuang, and Yonghong Tian.2016.Unsupervised Cross-Dataset Transfer Learning forPerson Reidentification.In IEEE Conference on Computer Vision and Pattern Recognition.1306-1315; H.Wang, S.Gong, and T.Xiang.2014.Unsupervised learning of generative topic saliency for person re-identification.British Machine VisionAssociation Bmva(2014)), the data introduced by this method Since the prerequisites of independent and identical distribution are not satisfied, a lot of errors are often introduced, and the performance improvement is very limited; the third is to obtain more data to expand the data set by linearly transforming the existing data set (see Niall Mclaughlin, Jesus Martinez Del Rincon, and Pau1 Miller.2015.Data-augmentation for reducing dataset bias in personreidentification.In IEEE International Conference on Advanced Video and Signal Based Surveillance.1-6), this method cannot provide diverse augmented samples ,and The expanded data set is not robust enough and the implementation complexity is too high

Images