Video segmentation and stabilization for BallCam

We present a video stabilization algorithm for ball camera systems that undergo extreme egomotion during sports play. In particular, we focus on the BallCam system which is an American football embedded with an action camera at the tip of the ball. We propose an activity-aware video stabilization algorithm which is able to understand the current activity of the BallCam, which uses estimated activity labels to inform a robust video stabilization algorithm. Activity recognition is performed with a deep convolutional neural network, which uses optical flow.