Ice-hockey puck tracking is a non-trivial task in hockey video analysis as it highlights the puck in the video for broadcasting, tactical play analysis, or referee assisting. However, difficulties, such as high speed, low texture features in images and constantly changing shape, make well-developed object tracker fail to track the puck. This paper introduces a real-time online-learning ice hockey puck detection and tracking system solely depending on video input to tackle this problem. The proposed approach categorizes pucks into free-moving and control-moving states, using a combination of contour fitting, correlation filter, and motion estimation techniques to detect and track them. A thorough analysis is performed focusing on the tracking scenario using broadcast video. To our knowledge, this is the first approach addressing detection and tracking nearly invisible high-speed pucks when shooting actions take place. Experiments with a comparison between a previous work targeting puck tracking show promising results in detection and tracking the ice hockey puck through broadcast video.