3D Reconstruction of Sport Scenes
Nowadays, the use of applications involving image analysis, geometric modeling or computer vision is common in transmitting sporting events. These applications allow, for example, annotation of sequences or the automatic selection of highlights of an event for a posteriori access [EKI 03, REN 09, YAN 08]. However, applications for three-dimensional (3D) reconstruction of sporting scenes remain extremely rare. This is due to the quantity of scientific issues caused by highly variable capture conditions (lighting, field, player kits, etc.) and the absence of information concerning acquisition parameters, leading to the use of a “raw” video flow (unknown camera parameters). Furthermore, the use of these applications during live events can only be envisaged if reconstruction is possible in real time (or almost real time). Complex systems have been proposed to overcome these difficulties, involving the use of multiple dedicated cameras; these systems represent a considerable outlay, can be difficult to implement and do not produce the desired automatic reconstruction [AMI 13, ORA 13, RED 13]. Certain research projects have focused on the automation of the reconstruction process using uncalibrated cameras and have offered innovative solutions [GRA 07, KIL 07, LE 06, LLU 11, TRI 13].
In this chapter, we will present the key stages of the reconstruction process, focusing on certain difficult points. In section 21.2, we will ...