Changing the Game: Novel Computer Vision Algorithms for Automated Sports Analytics

我们的目标是对体育视频数据中的人类活动进行全面的标注。我们的长期目标是自动化的方法，可以确定每个人在每个时刻的动作，标记玩家之间的互动，并对社会角色和更高级别的游戏情况进行分类。我们将为这一长期目标开发新的模型和学习算法。 我们的短期目标是开发能够处理领域中不同结构输入的深度结构化学习框架，包括跟踪和高级活动分析。我们将学习玩家检测、跟踪和活动识别任务的复杂深层表示，包括每个玩家执行的原子操作以及玩家与玩家的关系。此外，我们的目标是开发在线、资源受限的跟踪算法。 学习每个任务的独立表示是次优的，除非可以访问每个任务的非常大的训练集。为了减少所需的标签数据量，我们建议开发新的学习算法，该算法可以学习深层表示，同时考虑到所有这些任务的依赖关系。这将允许每个任务利用提供给其他任务的监督。 由此产生的算法和学习方法不仅将推动快速增长的运动分析行业，无论是专业运动还是业余运动，而且还将有助于从视频理解人类活动的进步。

We aim for a comprehensive annotation of human activity in sports video data. Our long-term goal is automatic methods that can determine actions for every person at every time instant, label interactions between players, and classify social roles and higher-level game situations. We will develop novel models and learning algorithms towards this long-term goal. Our short-term objectives are to develop deep structured learning frameworks capable of handling varied structural inputs in domains including tracking and high-level activity analysis. We will learn complex deep representations for the tasks of player detection, tracking, and activity recognition including atomic actions perform by each player as well as player-player relationships. Further, our goal is to develop online, resource-bounded tracking algorithms. Learning independent representations for each task is sub-optimal unless one has access to very large training sets for each tasks. To reduce the required amount of labeled data, we proposed to develop new learning algorithms that can learn deep representations while taking into account the dependencies of all these tasks. This will allow each task to leverage the supervision provided to the other tasks. The resulting algorithms and learning methods will advance not only the rapidly growing sports analytics industry, both for professional and amateur sports, but will also contributed to progress in human activity understanding from video.