Background-modelling techniques for foreground detection and Tracking using Gaussian Mixture Model

Abstract

Background Modelling and Foreground detection in sports has been achieved by cleverly developing a model of a background from a video by deducing knowledge from frames and comparing this model to every subsequent frame and subtracting the background region from it, hence leaving the foreground detected. This output from GMM background subtraction is fed into the feature extraction algorithm, which segregates the players based on teams. By extracting information of primary colors from each frame, the design of the algorithm based on the color of preference is done. Tracking algorithms Kalman and extended Kalman Filters help to predict and correct the location of players and in correctly estimating their trajectory on the field. Challenges such as shadowing, occlusions and illumination changes are addressed. The designed algorithms are tested against a set of performance parameters for the following datasets (Norway and FIFA) using MATLAB (2017b) and the inferences are respectively made. Object detection, motion detection and Kalman filter algorithms are implemented and the observed results are 100%, 84% and 100% accuracy respectively. With the results quantification and performance analysis, it is observed that with the decrease in contrast between player jerseys a decrease in detection accuracy occurs and with players crowded regions on the field and occluded players a decrease in tracking accuracy was observed.