Fast unstructured road detection and tracking from monocular video

In this paper, a fast particle filer based unstructured road detection and tracking algorithm is presented. We take the parameters of the road model and the relative pose of the vehicle with respect to the road as the state vector. The pixels of the test image are classified by learned boosted classifier based on rich pixel features to get a probabilistic output. For each particle, the virtual road image in the perspective view is generated according to the state vector. The particles are then weighted by the consistency of the virtual road image with the probability map. Then we can can estimate the optimal state with the particle weights. We further propose a scheme to accelerate the algorithm substantially with little degeneracy in performance by measuring the consistency with only several rows instead of the whole image. Extensive experiments show that the proposed method can detect and track the road robustly in various unstructured environments within real time.