Cricket Stroke Recognition Using Computer Vision Methods

The pressures on athletes to perform well in the sporting domain has drastically increased over the last decade and recent incidences of cheating and doping contribute to this. To further improve the game, computer vision methods are used to assist in certain tasks, such as the detection of no-balls, which reduces the number of unfair dismissals of a batsman. This research paper proposes a cricket stroke recognition model that can also contribute within the cricketing domain. Four different classes of strokes are analyzed namely: block, cut, drive and glance. The use of each stroke is dependent on the positional placement of the batsman’s head, feet, bat and hands. Images that exhibit these qualities are compiled and applied across three different machine learning algorithms to establish which algorithm is most applicable for the cricketing domain. Before images can be fed into the algorithms, each image needs to undergo preprocessing techniques to remove noise from the images and to make it easier for the subsequent algorithms to derive the stroke recognition. To remove noise each image undergoes a 7 × 7 kernel blurring process, followed by dilation, which highlights the foreground picture. Lastly, the histogram of orientated gradients is extracted and placed into feature vectors for each image. The feature vectors are then classified using K-Nearest neighbors and support vector machines. Additionally, another type of pipeline is employed that uses Convolutional Neural Networks (CNNs), AlexNet architecture. The AlexNet algorithm produced the highest model accuracy of 74.33%, which is further supported through the use of metric comparisons. It is realized that the AlexNet algorithm establishes a greater balance between precision and recall, thus predicting fewer false positives. The comparison between the algorithms and previous works show there is value in exploring newer computer vision methods for the use case and contributions toward the cricketing domain.