In order to make the program, we have to separate the program into different parts first. We will first have to find the necessary data for the different techniques and movements of fighters to train the model. Afterwards, we have to create the technique analysis AI model to identify the different types of punches used in the match and the vulnerabilities and mistakes made. 

This will consist of using the pre-existing database we found and training the model (MobileNetV2) using the training test split from the database. We will then evaluate it using the test data split from the pre-existing database found. The model will then be finetuned accordingly until it is able to correctly identify the techniques used and the possible vulnerabilities the players show during their fights (e.g. droppings hands, lowering guard unnecessarily, leaving body too exposed) with an accuracy of >90%. 

We will also have to create the player tracking and analysis AI model to be able to identify and track the player’s movement. We will create it using YOLO (you only look once) object detection model, importing the pre-existing database of images with annotations on players. These annotations allow for them to work as classes for YOLO, allowing YOLO to be able to track the players accordingly. We will train the AI model using the training test split from the database, and evaluate it using the test data split from the pre-existing database found. The model will then be finetuned accordingly until it is able to track the player’s movement  with an accuracy of >90%. 

We will also be using optical flow to track the glove’s movement. This is because the AI model will lose track of the glove movements due to very blur frames which happens a lot due to the punches being thrown very fast.

After we have both AI models, we will be able to detect the player, each of their movements, the accuracy of their punches and movements and possible vulnerabilities shown during the fights. These will all be displayed to the user in real time when using the program, with the vulnerabilities having their own timestamp corresponding to when the players show vulnerability in the video. 

Our product will also display the overall statistics of each fighters using the custom trained AI models predictions. This gives coaches a comprehensive view of their opponents tendencies and preferred techniques, and allows them to develop counter-strategies tailored to the opponent, enhancing the boxers’s preparation. Additionally, it gives them a more detailed insight on their own boxers tendencies, allowing the coach to refine their training to enhance positive aspects of the boxers and systematically work on correcting any the negative aspects.

In order to process a 2-minute video in under 10 minutes, we will use frame sampling techniques to reduce the frame rate without compromising the accuracy of the analysis too much. This will allow us to shorten the length of the video and process it faster, allowing us to process a 2 minute video in under 10 minutes.

Developed by: Jayden, Qi Jun and Tze Raye