About The Project
The goal of this project is to develop an interactive annotation system for locating and tracking players with their identities in broadcast sport video sequences. Moreover, the system is designed to allow users to add and correct errors caused by automatic predictions interactively. In fact, our system combines computer vision and machine learning algorithms to automate the tracking, detection, localization and identification procedures. First, it detects players by learning their shapes and tracks them based on their motion. Second, the play field is registered by the homography in each image in order to locate the foot position of players on the field. Finally, the appearance of each player is learned from automatically generated tracking results and propagates the information across all tracks of players based on conditional random field.
We have collected several broadcast footages of hockey and basketball games, where each of them is captured by a single broadcast camera that pans, tilts and zoom in/out, and test our system over them. The system was successfully tested on over 35,000 images of hockey data, equivalent to one period of the NHL game, and over 100,000 images of basketball data, equivalent to an entire NBA basketball game.
The "A Boosted Particle Filter: Multitarget Detection and Tracking" paper won the "Best Paper" prize in Cognitive Vision in the European Conference on Computer Vision(ECCV). [Slides]Demo