An interesting and broad field of application for machine learning is in the sports scenario. For most sports, an AI algorithm can take place either for athletes preparation or to assist in the scoring such as the VAR (Video Assistant Referee) in soccer.
In past releases, RidgeRun has worked on the action recognition for sports, more specifically in the action recognition for Mixed Martial Arts towards an automatic scoring system. As with most deep learning applications, the data preparation is one of the key steps for the correct training and inference. In the case of MMA automated scoring, the network requires to consume video frames with bounding boxes locating the position of each fighter, so an action can be estimated from it. Although trivial for a human being, it is not for a computer system, where a simple person detection algorithm will provide erroneous bounding boxes from people in the audience or referee instead of the fighters.
With this problem in mind, we ran a research on state-of-the-art algorithms in object detection and trained them for MMA fighters detection in video. In our report, we provide details on the creation of the dataset, the selection of the different models, the different training phases, and a final analysis towards processing time and footprint towards its execution on a NVIDIA Jetson Nano SoC.
The resulting model was ported to work with NVIDIA Deepstream and achieved 41 fps with a IOU of 0.44 for a 1080p video source. This result proved the model to be fairly successful in achieving the detection required for posterior action recognition stages.
As stated earlier, this effort is just a small part of the applicability of machine learning algorithms on sports and opens the door to translate it to other applications, not only MMA. If you feel interested in reading more about this research, you can look at the summary report at our developer’s wiki page.