Research And Implementation Of Person Re-Identification In Natural Scene

Person re-identification is an important technology in the field of computer vision.It can identify and track different pedestrians in the cross?domain monitoring system.It has important practical significance for criminal investigation,traffic analysis,and searching for lost elderly and children.In recent years,person re-identification technology has received extensive attention,especially with the application of convolutional neural networks.It has greatly promoted the development of end-to-end person re-recognition,and many research categories have emerged.However,in practical applications,person re-identification tasks also face some difficulties and challenges,such as the lack of a large amount of training data.The current collected data is very limited compared with the spatial-temporal distribution of real data.The visual characteristics of persons will change dramatically because of changes in background,illumination,posture,etc.,which will increase the difficulty of matching.How to overcome the impact of the above factors is the key to solving the person re-identification technology.This paper is based on the monitoring video data collected under natural scenes.Through the analysis and comparison of traditional and deep learning network methods,the overall research has a clearer understanding.This paper focuses on the Generative Adversarial Networks,which has the ability to generate samples,and has been widely used in the field of person re-identification.Aiming at the season differences in datasets and the shortage of training samples,this paper studies image style transfer methods based on the cycle-consistent adversarial networks and proposes a new solution.An additional constraint module is used to ensure that the generated samples are more in line with the person re-identification application scenario.Experimental results show that this method can effectively shorten the season difference between datasets.This paper produced a new multi-season person re-identification dataset Traffic4340 as a benchmark for experiment.Compared with the current public datasets,Traffic4340 collects person images under the natural traffic cameras,which not only has winter data but also summer data,which is more in line with the real application scenario.Based on this dataset,comparative experiments can also be conducted to analyze the state-of-the-art works.