In this work we present a method to automatically learn and detect cast shadows on highway surveillance scenarios. The first stage of this method uses a weak classifieras a pre-filter to select possible shadowed pixels in order to learn multi-layered statistical shadow models using a recursive Bayesian learning approach. These models will then be used, by a strong classifier, to correctly distinguish shadows. To prevent misclassifications from corrupting the results of both classifiers, spatial and temporal dependencies are also taken into account. Based on a Conditional Random Field (CRF), spatial/temporal dependencies in traffic scenes are formulated under a probabilistic discriminative framework, where contextual constraints during the detection process can be adaptively adjusted in terms of data-dependent neighborhood interaction. To meet real-time requirements the CRF energy function is minimized using the Dynamic Graph Cut algorithm (DGC) for the st-mincut/maxflow problem. This fast and fully dynamic algorithm uses the solution of the previous graph cut computation for solving the new instance of the problem. Our solution is data-driven and recursively adapts the shadow models to the changing conditions. The accuracy improves as it processes frames that contain shadows and is not limited by preset values. This technique is being used in a real outdoor traffic surveillance system in order to minimize the effects of cast vehicle shadows. Experimental results demonstrate the effectiveness of the proposed framework and the advantages of applying spatial contextualization and temporal coherence to the weak/strong classifiers.
展开▼
