The Research Of Ice Hockey Player Temporal Spatial Data Extraction Technology Based On Computer Vision

Spatial temporal data driven performance analysis in competitive team sports,has been widely developed in soccer and basketball,is an effective approach to improve the training level and enhance the competitive capacity.However,high physical intensity and frequent substitution in ice hockey games increase the difficulty of extracting player tracking data,which restrict the development of related researches.In order to extract ice hockey player spatial temporal data from the broadcasting videos,this paper proposed a spatial-temporal data extraction technology based on the understanding of ice hockey players appearance.The proposed technology includes three parts:spatial player detection,team affiliation detection of detected player and temporal player trajectory reconstruction.The specifical process is as follows:(1)Dataset:We collected and labeled two ice hockey game datasets:The Winter Olympic ice Hockey dataset and the National Hockey League(NHL)dataset;(2)An appearance-adapted two-stage cascaded Convolution neural network(CNN)model was proposed to detect ice hockey players in game videos,where stage one network was designed to exclude most of the background regions and stage two network handled the harder samples that cheated the first stage;(3)Based on the player detection results,a simple but effective team affiliation detection method was proposed subsequently to recognize players,which made use of the center areas of player detections to extract uniform color features to determine the team affiliations;(4)Based on the player detection results,an IOU calculation based trajectory reconstruction method was applied to connect detection bounding boxes in different frames that referred to the same player to produce a player trajectory.The results of experiments showed that the proposed method performed well on spatial-temporal data extraction of ice hockey player:the precisions of player detection and team affiliation detection were 98%and 93%,respectively,the average frames of trajectory reconstruction is about 25%of the whole sequences.Details are as follows:(1)The player detection experiment conducted on The Winter Olympic Game ice hockey game dataset showed that the proposed player detection method achieved the mean accuracy of 98.75%,the mean recall rate of 94.11%and F-score above 95%.In the comparison experiments with other state-of-the-art methods,the proposed method was also competitive which outperformed Yolov3 and CornerNet but underperformed Faster-RCNN method.The comparison experiment conducted with the Baseline model showed that the adjustment of critical parameters improved the performance significantly.(2)In the player affiliation detection experiment,the proposed method achieved the mean accuracy of 93.05%in four different games,which was able to recognize players belonged to different teams with extra annotations.(3)The player trajectory reconstruction experiment was conducted on two sequences of NHL games,which were a power play sequence and a fast break sequence,respectively.The proposed method performed greatly on trajectory reconstruction with mean tracking frame lengths of 77 and 31 in consideration of trajectory breakage caused by empty detection input.Above all,the proposed spatial-temporal data extraction technology of ice hockey player could effectively extract the positions of players and provide dynamic information of player positions for tactical analysis.