Recognizing vehicle-contours with a compositional deformable model

This paper illustrates a compositional deformable model for detecting vehicle and recognizing vehicle-contours. To overcome the difficulties that vehicles in an image have various sizes, shapes, colors and poses, this model has two main characteristics: first, the model is made up of constituent parts which shared by vehicles. The locality of parts give the model the ability to recognize vehicles with different types (e.g., although vehicles have various sizes and shapes, they are usually composed by roof, windscreen, windows, etc.). Second, the spatial relationships of these parts are represented by Markov Random Field (MRF). The model is deformable to adapt to vehicles of different shapes and poses because of the appropriately changing of combinations of these parts in the MRF. Experimental results with real world images show that this method is effective in vehicle detection and vehicle-contours recognition.