State-of-the-art object detection model for detecting CCTV and video surveillance cameras from images and videos

Abstract

As the current GDPR law in the EU prohibits unnecessary use of CCTV cameras in public places, and privacy concerns of smart CCTV cameras have been raised, CCTV cameras cannot be regarded as just easy tools to help secure your assets. Smart technologies such as facial recognition have reached CCTV cameras and the amount of data gathered from them is expanding rapidly. This thesis explores how to use state-of-the-art object detection architectures and frameworks to create models to detect CCTV cameras from images (e.g., street view) and video. A literature review was performed to establish a fundamental understanding of object detector algorithms. The metrics from Microsoft Common Objects in Context were used to evaluate the detectors as state-of-the-art since most recent architectures have been tested with it. The selection process also took into account the framework around the detector as the tools in use are essential for the future refinement and adoption of the model. Three detectors were identified as prime candidates; CenterMask, ATSS, and TridentNet. This thesis developed a set of state-of-the-art detectors for 'CCTV-camera' objects achieving over 90% mAP@0.5 and with further possibilities, improvements, and testing datasets being disclosed. This thesis is a part of a three-way project collaboration with two other M.Sc. theses being written by Tuomo Lahtinen and Lauri Sintonen. The intention is to create a toolset to improve the detection model further and to use it for mapping CCTV cameras from street view images and create safety-centric routing and navigational suggestions.