Sports Activity Recognition and Classification

Conventional Search engines search for results using keywords. These search engines matches the contents with the keyword and if they find match the provide us the results. But they cannot do it by just observing the content. Using machine learning and image classification a system can provide us the results which were produced by only observing the image or video not the embedded texts. Image classification is a powerful machine learning based application of deep learning. Because deep learning techniques can perform effectively when implemented on large scale image data they are becoming more popular day by day. In this research, we tried to create a system with modified classification model which can detect what type of sports is being played in an image or a video effectively using our own dataset. Our model can classify images which contain almost identical sports being played. As it is not feasible to solve this kind of problem using classical machine learning models we implemented deep learning to solve our problem. We used transfer learning as this technique is more efficient and gives more accuracy. We implemented MobileNet[1], MobileNetV2[2] and ResNet[3] as base models for our modified model. Out of these the modified model with MobileNet as base model gave us the accuracy of 81.89%
Department Name
Electrical and Computer Engineering
North South University
Printed Thesis