Modeling the Effect of Relative Camera Pose Differences on Change Detection Tasks

Rafal Muszynski, Hiep Luong

Abstract: Change detection algorithms have to be robust to unrelated changes in the observed scene such as changing lighting conditions or shot noise. When the camera is not static, image registration is often performed as a preprocessing step. Unfortunately, image resampling performed during registration can introduce additional artifacts which are a source of false positive detections. We present an approach to modeling changes resulting from observing the scene from a different pose. We reformulate the general problem to only moving the camera further away from the target object and model the effect of observing the scene from the new viewpoint using the camera's point spread function (PSF). Our proposed model significantly outperforms the baseline approach and we are able to characterize the performance of change detection algorithms more reliably. The proposed model can be applied to guide the user in acquiring good quality images at runtime for change detection or to design constraints on the camera pose to ensure required accuracy of change detection algorithms.