Real-time Head Pose Estimation inLow-resolution Football Footage UsingRandom Forests

Abstract: This report presents a method for real-time head pose estimation in low resolution football footage. The presented method uses a random forest trained on synthetically generated head images. The use of synthetic training images is shown to be a good substitute for the use of manually labelled images. The presented method compares favourably to support vector machines trained for the same task, both in terms of accuracy and speed. It is noted that the method relies on a good head detection to perform well. The report also examines ways of combining the image based head pose estimation with contextual features such as ball position and player position. It is shown that the relative direction to the ball can improve the accuracy of the pose estimate in certain situations. Furthermore, it is found that the random forest method can easily be extended to incorporate images from multiple cameras to improve accuracy.