Multi-view Pedestriain Re-identification Algorithm Research And Data Collection

As an important part of public security,video surveillance technology has gain rapid development in recent years.Thanks to the decrease in the costs of surveillance devices and the increase of their quality,public areas covered by surveillance cameras grew,rendering the traditional ways of human watching security screens to track and identify pedestrians futile.As a result,intelligent pedestrian video surveillance problem is becoming a hot interdisciplinary research topic,with major influence in computer vision area.The goal is to use computers to analyze security footages,in order to identify the locations where pedestrians appear,and then match and track identical persons across long distances and different scenes,in effort to reduce the human power required for surveillance work while improving accuracy.However,as the population rapidly grows,some urban public areas can get very crowded during rush hours,which,when observed by security cameras,can result in people being obstructed in the view by scenery objects or other people.These scenarios,which are difficult to be handled automatically by machine vision,bring great challenges to the designing of large-scale intelligent pedestrian surveillance solutions.The thesis focus on the research topic of intelligent pedestrian surveillance.First,we introduce how the problem in academic research is divided into several basic sub-problems such as human detection,tracking and re-identification,and how the raw video data passes through these parts while information is extracted and analyzed.As a part of this,some major related works on these sub-problems are overviewed and introduced.Additionally,we analyzed in detail the primary topic of this thesis,human re-identification,specifically on its position in the intelligent pedestrian surveillance problem framework and the significance of solving it.Then,we further discuss the re-identification problem by researching and reviewing the existent work,in order to discover the main research focuses,which are extracting more robust features from human images and learning of better feature distance metric.Based on the learning of prior works,a novel pedestrian recognition method based on deep learning,more specifically convolutional neural network(CNN),is proposed,which makes use of existing image classification network structure to extract and select features from input pedestrian images,and ultimately achieve better results in the CMC evaluation over traditional methods.In the next part,we focus on the datasets used in the research area of pedestrian surveillance.The first step is to review the major existing datasets used by various sub-problems of this topic,by analyzing the type of contents and ground truth provided by these datasets,and the benchmarks and evaluation protocols they designed for the according sub-problems.Then,aiming at the characteristics and shortcomings of the current datasets,we present a system for collecting pedestrian data on a large scale,including selecting the appropriate data capturing,controlling and auxiliary positioning hardware devices,and designing the collection node,software system and controlling network.Then we show the whole process of collection,the contents and formats of data collected,and the marking process afterwards according to the academic research requirements.