Car detection using Markov random fields on geometric features

Car detection is widely used in many vehicle management systems. Most of the car detection methods are focused on the top (aerial) or side images. Usually these methods are supervised detection which is complicated and need to use different kinds of classifiers and training data sets. This paper introduces a new unsupervised approach to provide fast car detection from rear or head images of the cars running on the road. The contribution of this paper lies in the following: (1) The car detection model is based on the blocks of pixels with consistent color/gray-level, instead of individual pixel. (2) Geometric features are extracted to represent car. (3) By retrieving the detecting car positions, the graph model is divided into several sub-graphs to compress solution space. (4) A Markov random field is defined on each sub-graph to restore the car rear or head shape by their inner and inter relationship. In this way, the car bodies with different colors in low resolution images can be fast detected from a complex background without any training data.