CASRM: Cricket Automation and Stroke Recognition Model Using OpenPose

With the rapid changes within sport, specifically cricket, technology has been used to cater to the challenges faced within the domain. However, research within the field of study has shown that there is a gap to bridge in the way of establishing a cost-effective means to recognize different cricketing strokes. In our previous work, feature extraction methods such as Histogram of orientated gradients with support vector machines, K-nearest neighbor, and the AlexNet architecture were used to achieve cricket stroke recognition. While promising results were obtained, this article will attempt to exploit OpenPose skeleton keypoints, which will be used as a set of descriptive features that will be fed into the Long Short-Time Memory architecture for cricket stroke recognition. By applying the OpenPose skeleton to the dataset, the model can capture the pose keypoints of the cricket batsmen, whereby the body part locations and detection confidence are presented as a feature vector. The image dataset, which was compiled in a previous study, is used to ensure a fair measure of the proposed model. The strokes that will be addressed are as follows: block, cut, drive and glance. The Long Short-Time Memory architecture outperformed previously tested classifiers with a recorded model accuracy of 81.25%. The results suggest the model is capable of recognizing different cricket strokes. As a result, a human-computer interaction system can be developed to assist coaches and spectators to gain further understanding within the domain.