Within a probabilistic framework, based on principal component analysis neural networks, a novel Monte Carlo tracking technique is suggested for autonomous navigation of land vehicles on unstructured roads. The straight road model is used to represent road edges and its status is predicted by the second-order autoregressive model. Color information and local spatial features are combined in the HSV color space in order to obtain principal components by principal component analysis neural networks. The status of the road is estimated by particle filters according to statistical features of road edge windows. Experimental results show that the unstructured road tracking can be robustly realized by the method.