Adaptive Method for Segmentation of Vehicles through Local Threshold in the Gaussian Mixture Model

Abstract

The segmentation of vehicles is a non-linear problem that has been tackled using methods for background subtraction in systems for traffic control. Probabilistic models, such as Gaussian Mixture Models (GMM), estimate the background of dynamic environments in this approach. The general modeling considers independent distributions for each pixel of the image. So, the classification is performed singly. The system uses often only one threshold to classify the pixels into background and foreground regions. This approach doest not work well when the cluster intersection is significant. In the vehicle segmentation, the color of the vehicles are similar to background, so the accuracy is affected. This paper proposes an approach to improve the classification of traffic scenes. This approach uses local thresholds to encourage the segmentation of vehicle regions. These thresholds are estimated by a spatial analysis of the previous classification. The results of the experiment performed shown that the classification process is improved by this approach.