Cricket Stroke Activity Recognition in Telecast Videos

Abstract

Activity recognition in videos finds application in surveillance systems, elderly care, workflow monitoring, content-based searching, etc. With the increasing amount of video data, there is a huge scope for models that accurately recognize the actors and actions occurring in the videos. A major application area, where there are large volumes of video data already available and more is being generated daily, is sports activity understanding. Sports activity recognition in videos is quite challenging, mainly, due to the presence of camera motion induced noise, multiple camera viewpoints, frequent camera switching, multiple actors performing complex actions, and unavailability of domain-specific annotations. Recognizing sports activities usually involves spatio-temporal detection and tracking of players, objects and events of interest followed by modeling the interactions between them. Some of the recognition and tracking approaches applied to a specific domain, for example Soccer, may not be generalizable to other sporting domains, like Baseball. Each individual sporting domain needs some domain-specific assumptions which can be used for efficient modeling of the recognition problems and tasks. newlineWe consider Cricket telecast videos and solve the tasks of temporal localization of Cricket strokes, their recognition based on direction of motion (coarse-grained) and the type of cricketing shot played by the batsman (fine-grained). newline newline