The purpose of this project is to showcase how ChatGPT can serve as a valuable programming aid, particularly when working with unfamiliar libraries. The videos are quite long and are meant so that someone could follow along and reproduce the process if so desired.

Here is a link to the Google Drive folder containing the data files.

I utilized ChatGPT to demonstrate how it can effectively handle the majority of coding tasks in a Google Colab notebook that analyzes a Fencing Action and acts as a fencing referee to determine touches. To accomplish this, I divided the project into three main sections: Data Collection, Detection and Tracking, and Model Training.

For Data Collection, I leveraged footage from the Red Strip captured during Day 2 of the European Fencing Championships. Using the overlay from the video, I detected instances when touches were scored and extracted the corresponding scores. Next, I sorted the Epee and Saber clips based on scores resulting in a point and when the clock read 3:00 minutes.

Using the clips obtained during Data Collection, I employed YoloV8 by Ultralytics in conjunction with Roboflow to assist with dataset management and create a customized detection model. Subsequently, I utilized this model, along with basic tracking analysis, to trace the paths of the fencers' bellguards. By analyzing the bellguard positions, I derived the acceleration of the bellguards and combined the bellguard x_position acceleration with the frames when the colored lights turned on. This information was then saved in comma-separated values (CSV) files. Additionally, I appended the touch data for each clip, which was obtained from the overlay analysis conducted earlier.

Following the recommendations of ChatGPT, I employed a Gate Recurrent Unit (GRU) model to process the collected data. Due to the simplicity of the data and a relatively small number of inputs, the model trained quickly. Subsequently, I applied the trained model to analyze Clips 1, 2, and 3, which were extracted from the Blue Piste footage captured during the same event.

By following this approach, I demonstrated how ChatGPT can be a valuable asset in automating various aspects of a fencing analysis workflow.