Fast and Reliable Detection of Hockey Players

Current popularity of augmented reality (AR) stems from its ability to enhance the perceived environment in real- time with additional information of semantic context, such as sports scores shown on TV during match broadcasting. Its other application areas range from industry and medicine to military, commerce and entertainment. Advanced AR technologies obviously rely on accurate, yet fast enough algorithms for multimedia processing and object recognition. In this paper, we will study the possibility of using convolutional neural networks (CNNs) for real-time detection of hockey players from video streams of broadcasted ice-hockey matches. Supporting experiments performed so far yield sufficient accuracy for this task (above 98.5%), while maintaining reasonable computational demands and acceptable robustness both with regard to noise and minor image transformations like translation, rotation and scaling.