Object detection in infrared image with fast nonparametric background model

A background subtraction method is introduced with nonparametric background model for infrared surveillance application. This model employs a sample set as the statistical model of each pixel, and calculates conforming possibility of a pixel's value with kernel estimation. Two thresholds are adopted for object detection and model updating, which segments the frame into three categories: reliable background, unreliable background and interest region. Interest region is segmented into intruding object and false positive detection with context provided by unreliable background. Experiments with several infrared image sequences show that this method could precisely detect salient intruding object and weak intruding object that is easy to be confused with noise.