Boxing Punch Detection Using Computer Vision

Mark McQuade shares his experience of using computer vision to develop a model that detects and classifies different types of boxing punches, such as jabs, crosses, hooks, and uppercuts, during a workout session. Initially, the project involved capturing data from a 12-minute boxing session and annotating 239 images, which were then processed using the Roboflow platform to create a dataset. The model underwent several iterations, including the addition of a 'no punch' class and refinement of bounding boxes to improve accuracy. Despite initial challenges, the third version of the model performed better in distinguishing punches, although it still struggled with differentiating similar punch types like hooks and crosses. Further improvements could be made by increasing the dataset size and incorporating more pre-processing techniques. The project culminated in using Roboflow's video inference tool to analyze an entire boxing round, with eventual plans to calculate punch statistics to enhance workout effectiveness.