Closed-Form Change Detection from Moving Light Field Cameras

We present a novel approach to change detection from a moving camera, a task normally complicated by the nonuniform apparent motion of complex 3D environments. By combining methods for visual odometry and rendering from light fields, we derive what is, to our knowledge, the first closedform approach to 3D change detection. No depth estimation is required – apparent motion is disregarded by exploiting the geometry implicitly encoded by the light field. The method has constant runtime and is suitable for real-time implementation on parallel hardware including FPGA and GPU. We show qualitative and quantitative results for Lytro-captured imagery, with our method significantly outperforming a naive 2D per-pixel method. We discuss a generalization of the derivation to other traditionally still-camera-only methods.