Spatio-temporal Localization of Robbers in Videos

Abstract

Robbery -- the use or threat of force to take another's property -- is among the nation's, most serious crime problems. High rates of robbery plague many inner-city neighborhoods. Robbery is a central component of the fear of crime. Many suburban residents, whose objective risk of victimization is much lower than that of their inner-city counterparts, nevertheless manifest substantial 'fear of being robbed. This has important consequences for personal freedom; fear of violent victimization can lead people to a limit their public activities. The purpose of this project is to find a solution which is efficient enough in detecting and localizing robbers. This would have a lot of significance in places where the victims are helpless and cannot contact the authorities. The project using advance machine learning techniques to detect and localize robbers seen from the CCTV footage. Using just the camera screen, we are able to not only detect but also fully localize robbers in time and space. Using Action Tubelet detector we create multiple tubes where the robbers and identified and after regressing the tubes we can fully localize robbers. Although this project is very meaningful in our society, it carries a lot of hardware limitation. For this system to work in real time, we require computationally expensive GPUs which are very costly.