Research On Vehicle And Pedestrian Retrieval In Video

Nowadays,monitoring system is everywhere in city because of the development of the science and technology,which generates the large amount of the video.Therefore,how to find one specific pedestrian or vehicle across the multi-cameras has a significant meaning in monitoring system.This paper researches othe problem of the re-identification of the vehicles and pedestrians across the multi-cameras and designs related method.The object we study appears in videos,so besides the object re-identification,our work also includes object detection.The frame of traditional object re-identification in videos consists of object detection and object re-identification.To improve the performance of the method,we add object tracking to the frame so that the information of the query object can be more specific.What’s more,we collect the object re-identification results’time and position information in the videos,and propose a concept called local average of similarity rank,which can predict when the query object appears in the videos.Then according to the object’s occurrence time and location,we optimize the result.In this paper,we adopted object detection and object re-identification method based on deep learning in our method.To prove the superiority of the deep learning,we did experiment both by traditional object detection method and deep learning object detection method.In the part of re-identification,we design two re-identification solution for pedestrian and vehicle and use modified convolutional neural networks feature to describe the pedestrian and vehicle.We use PKU-SVD-B database to conduct the experiment,in the part of pedestrian re-identification in videos we use the pedestrian tracking across cameras of the PKU-SVD-B dataset and in the part of vehicle re-identification we use the accurate vehicle re-identification of the PLI-SVD-B dataset,and we also compare our methods with others traditional methods...