DEEP LEARNING MODEL FOR 2D TRACKING and 3D POSE TRACKING OF FOOTBALL PLAYERS

2D and 3D multiple object tracking, is an open problem inside the computer vision community with multiple applications in the industry such as in the autonomous vehicles or in the sport field. Many works have been conducted in the past to solve and improve this task, especially for person tracking due to its greater interest. Recently, the deep learning techniques have been able to beat the state of the art in tasks such as image classification, object detection or 3D pose estimation. Thus, this work has made use of deep learning methods to build a 2D and 3D tracking applications. These techniques are combined with a tracking by detection scheme to perform the tracking and achieve a good result. Contribution of work proposed in this thesis would be two-fold. First, it is implemented multi-person object tracker in 2D which specialized for sport like soccer. Second, a 3D multi person tracker is designed, which inputs single RGB image and outputs the 3D poses with IDs.