| In the field of computer vision,action recognition in videos under intelligent has got much academic attention in the past few years.However,the scene types of video are abundant and miscellaneous,such as street surveillance video,indoor surveillance video,and competition video of a certain sports event,etc.It is more practical to recognize action in a video under a specific scene.For badminton singles video within sports event,in order to better assist coaches to analyze stroke of the major player and make users can enjoy the diversified demands of each common stroke meta-video set,this paper research the approach of stroke recognition in badminton videos based on Temporal Segment Networks.The main research contents and conclusions of this paper are as follows:(1)Highlights extraction from a badminton videoAmong the broadcasting view of a badminton video,the segments have relatively vigorous exercise of players and long enough playing duration are defined as badminton video highlights.In order to make users directly watch highlights of a badminton video to save their time,this paper research a method of obtain useful highlights by view classification and a way to screening out badminton video highlights through capturing players’ speed velocity based on target tracking.The effectiveness of highlights extraction is evaluated in experiment.The actual effective highlights accounted for 93.1% among the extracted highlights,indicate that the result of highlights extraction is favorable,which lays a foundation for making the dataset of badminton meta-videos.(2)Badminton stroke localization based on pose estimationIn view of the limitation that present methods can’t locate badminton strokes well,propose an algorithm of arm swinging amplitude based on pose estimation to locate the time domain of strokes in a highlight,and further implement meta videos extraction in a highlight.The experiment results show that the indicator Io U(Intersection over Union)of badminton stroke localization is 82.6%,indicate that the method proposed in this paper can effectively locate badmonton strokes in the highlights,which enhance the practical meaning of badminton stroke recognition.(3)Badminton stroke classification based on temporal segment networksIn this portion,training temporal segment networks which imbedded a lightweight attention module to get the classifier of badminton strokes.The classification results include four common types: forehand,backhand,overhead and lob.Meanwhile,if a meta video belongs to overhead,further divided it into clear or smash stroke by the proposed approach based on morphological processing.Experiment results show that the indicator AUC(Area Under Curve)of each badminton stroke predicted by classifier is above 0.98.The average recall and precision of the final five strokes(forehand,backhand,clear,smash,lob)respectively reach 91.2% and 91.6%,which indicates the classify model can effectively recognize a badminton stroke in a meta video.Finally,the badminton stroke recognition system is designed,which connects the badminton stroke localization program and stroke recognition program in series.With an extracted highlight as input data,the meta video collection of each stroke will be output end to end. |
