Improving motion-based object detection by incorporating object-specific knowledge

In this contribution, we describe an object detection method that jointly considers low-level features and higher-level object knowledge. The method partitions a stereo image sequence into its most prominent moving groups with similar 3-dimensional (3D) motion and of consistent object-specific appearance. Image segmentation is performed by a Bayesian Maximum a Posteriori estimator assigning the most probable motion profile to each image point. The motion profiles of the elaborated motion models are iteratively refined by an object tracking procedure. Additionally, the probability of salient points to belong to an object category is considered in the probabilistic framework. Our expectation on spatial continuity of objects is expressed in a Markov Random Field (MRF) model.