Pose Estimation of Players in Hockey Videos using Convolutional Neural Networks

Traditional hockey scouting procedures for evaluating player performance is based on visual monitoring of hockey videos and statistics. However, that evaluation is time consuming and prone to human bias. In addition, current research within hockey analytics quantifies player performances by employing statistical models on common hockey statistics. To improve statistical models and increase the precision of evaluations, gathering data that is specific to each individual player, as well as, evaluating a player based on the capability of a player’s performance is crucial. Techniques of skating, shooting, stick handling, passing, and speed of a player, are clues for identifying the performance of a player. In order to gather these key pieces of information from video data, the position of player limbs and joints should be found in each frame of the video. By knowing the location of joints in all video frames, one can easily track all body movements of players and use them for calculating desired information. In this work, we use a novel approach that determines a hockey player’s body placement in video frames, also known as pose estimation via a convolutional neural network-a computer vision algorithm that imitates the learning processing of a brain. The proposed method provides a tool to analyze the pose of a hockey player via broadcast video which aids in the eventual assessment of a hockey player’s speed, shooting etc., which, are alternatives to goal orientated statistics. The algorithm proves to be successful since it identifies on average 81.56% of the joints of a hockey player on a set of test images. Key wordshockey analytics, pose estimation, video analytics, convolutional neural networks, stacked hourglass network, compute vision.