Abstract:
"
Human Action Recognition (HAR) has become a significant domain since it is an efficient mechanism that
recognizes human activity. Though HAR has been utilized in several other fields, this research is focused
on the step detection of tennis strokes using HAR. Video analysis is an improved method to detect stroke
steps without involving any device which will interfere with the player or the racquet resulting in a change
in player performance. This method can be applied to video recordings recorded without any special
equipment, only using a smartphone or a conventional consumer camera. Identifying the stroke steps and
analyzing the performance is essential for a novice player and the coach.
This research demonstrates the individual step detection is a feasible task with a considerably large dataset.
Convolutional Neural network (CNN) and ImageNet VGG19 model architecture have been used to train
the models. The dataset was created by combining UCF 101, THETIS, YouTube videos, and manually
recorded videos since the lack of tennis videos is available. The batch processing approach has been utilized
to overcome storage problems. The rolling prediction averaging mechanism has been employed for reliable
and accurate prediction.
The output of the system is stroke type and stoke steps of specific tennis strokes. The tennis step prediction
ensures if the frames are in a sequence. Moreover, this research detects the steps of Serve, Forehand, and
Backhand used by the athletes consistently when playing tennis. The results indicate an accuracy of 66%
and 78% for stroke and step detection, respectively."
