ABSTRACT
Object detection is a critical preprocessing step in intelligent visual surveillance systems, and background modeling and subtraction is a natural technique for object detection in video captured by a static camera. We propose a pixel-based density-based background modeling and subtraction algorithm using multiple features, where generative and discriminative techniques are combined for classifi cation. In our algorithm, color and Haar-like features are integrated to handle spatiotemporal variations effectively for each pixel. A background model is trained for each feature independently with a Gaussian mixture based on kernel density approximation (KDA), where all the parameters for the Gaussian mixture are determined automatically and adaptively. Background subtraction is performed using a support vector machine (SVM) over the background likelihood vector for a set of features. The algorithm is robust to the spatial variations of background and shadow. We compare the performance of the algorithm with other densitybased background subtraction methods quantitatively and qualitatively with several challenging videos.
