Automated detection and tracking of slalom paddlers from broadcast image sequences using cascade classifiers and discriminative correlation filters

This paper addresses the problem of automatic detection and tracking of slalom paddlers through a long sequence of sports broadcast images comprised of persistent view changes. In this context, the task of visual object tracking is particularly challenging due to frequent shot transitions (i.e. camera switches), which violate the fundamental spatial continuity assumption used by most of the state-of-the-art object tracking algorithms. The problem is further compounded by significant variations in object location, shape and appearance in typical sports scenarios where the athletes often move rapidly. To overcome these challenges, we propose a Periodically Prior Regularised Discriminative Correlation Filters (PPRDCF) framework, which exploits recent successful Discriminative Correlation Filters (DCF) with a periodic regularisation by a prior that constitutes a rich discriminative cascade classifier. The PPRDCF framework reduces the corruption of positive samples during online learning of the correlation filters by negative training samples. Our framework detects rapid shot transitions to reinitialise the tracker. It successfully recovers the tracker when the location, view or scale of the object changes or the tracker drifts from the object. The PPRDCF also provides the race context by detection of the ordered course obstacles and their spatial relations to the paddler. Our framework robustly outputs the evidence base pre-requisite to derived race kinematics for analysis of performance. Experiments are performed on task-specific dataset containing Canoe/Kayak Slalom race image sequences with successful results obtained.