Background subtraction using spatial mixture of Gaussian model with dynamic shadow filtering

Many applications of computer vision, motion captures nowadays are an active research field. Supported by camera innovation in high definition technology and high-speed processing unit technology make higher degree on object detection standard. We can see it from the increasing number of new methods that have improvement in accuracy. In automatic vehicle surveillance area, Spatial Mixture Gaussian model becomes well-known moving based object detection via background subtraction technique in this decades. This method models particular pixel as mixture of Gaussians distribution with regard to pixel's higher probability of occurrences and variance of each Gaussians in the mixture model. Although, this model has threshold to control the sensitivity of object's motion, it has problem with separating an object from its shadow. This is happening because the shadow attaches to the object. Since they always move in tandem, as the result, detected object area will merge and shadow and object will form into a single unity that is difficult to separate. In accordance with detection, occluded object because of a shadow will decrease detector's accuracy. Therefore, we need to remove shadow, in order to maintain detector's quality of accuracy. Challenge in doing so is there is exist dynamic illumination condition which resulting a nonuniform shadow pixel value. This can cause failure of threshold-based linear shadow casting technique. To solve above-mentioned problem, we need a shadow filter that can adapt to the illumination changes. In this experiment, we have successfully implemented an adaptive shadow filter based on DSD algorithm to improve background subtraction method. Our proposed method has a stable result in outdoor environment dataset and it is proven to be able applied to traffic surveillance video application.