Person Re-identification and Intelligent Crowdsourcing with Applications in Public Safety

Video surveillance systems are of a great value for public safety. They are used as an effective tool for crime prevention and as an after the fact forensic tool. Many automatic methods have been proposed for video analytics such as anomaly detection and human activity recognition. Such methods have significant challenges due to object occlusions, shadows, various scales, and changes in viewpoints and illumination conditions. In addition, mobile or networked environments have limited bandwidths and adaptive data-rate streaming is frequently used. Video compression can introduce significant quality degradation that impacts the accuracy of video analytics. In this thesis, we propose a two-stage quality-adaptive convolutional neural network system for pedestrian detection to address the problem of a changing video data-rate. We also present a person re-identification method based on the use of a patch-based dynamic appearance model. Person re-identification is of interest to public safety as it helps law enforcement identify persons of interest when they re-appear in the surveillance system. We use deformable graph matching for person re-identification using histograms of color and texture as features of nodes. The method is evaluated using a dataset that was collected at the Greater Cleveland Regional Transit Authority. For forensic analysis of surveillance video, we describe an intelligent crowdsourcing system to utilize human intelligence and perform tasks that machines find difficult. We enhance crowdsourcing by incorporating computer vision methods and propose a hierarchical pyramid model to distinguish the crowd members based on their ability, experience and performance record