Sequential decomposition of 2D apparent motion fields based on low-rank and sparse approximation

Estimation of camera egomotion and detection/tracking of moving objects in video captured by moving camera are challenging tasks in many applications such as autonomous vehicle control, medical image analysis, video surveillance, compression, stabilization, etc. Camera egomotion and object motion in a scene both induce temporal variation of image intensity, which is represented as a vector field of the so-called optical flow. An egomotional optical-flow field at any moment is composed of a few common linear basis vector fields related to camera's translation and rotation, while the moving objects superpose spatio-temporally local and sparse optical flows. We present an online algorithm of stable principal component pursuit that decomposes frame by frame the apparent motion fields into linearly dependent optical-flow fields and sparse optical flows for a given sequence of images. We experimentally show that the former captures the egomotion, and the latter detects the moving objects.