Data driven stochastic approximation for change detection

Thomas Flynn, Olympia Hadjiliadis, Ioannis Stamos, Felisa J. Vázquez-Abad

Abstract: Online change detection has many applications, ranging from finance and manufacturing, to security and computer vision. Designing a change detector for use in a given domain can be very time consuming, and model-based algorithms often require knowledge of the underlying stochastic model. To address these issues, in this work we explore a supervised learning approach to a change detector. We implement a gradient based procedure to find the optimal parameters for a change detector. We demonstrate the methodology on both synthetic and real world data for classifying 3D laser range image data in real-time.