HyperStackNet: A Hyper Stacked Hourglass Deep Convolutional Neural Network Architecture for Joint Player and Stick Pose Estimation in Hockey

Abstract

Human pose estimation in ice hockey is one of the biggest challenges in computer vision-driven sports analytics, with a variety of difficulties such as bulky hockey wear, color similarity between ice rink and player jersey and the presence of additional sports equipment used by the players such as hockey sticks. As such, deep neural network architectures typically used for sports including baseball, soccer, and track and field perform poorly when applied to hockey. Inspired by the idea that the position of the hockey sticks can not only be useful for improving hockey player pose estimation but also can be used for assessing a player's performance, a novel HyperStackNet architecture has been designed and implemented for joint player and stick pose estimation. In addition to improving player pose estimation, the HyperStackNet architecture enables improved transfer learning from pre-trained stacked hourglass networks trained on a different domain. Experimental results demonstrate that when the HyperStackNet is trained to detect 18 different joint positions on a hockey player (including the hockey stick) the accuracy is 98.8% on the test dataset, thus demonstrating its efficacy for handling complex joint player and stick pose estimation from video.