This project leverages the power of YOLOv8 (You Only Look Once) and image segmentation techniques to track basketball shots and predict whether the ball will go in or not. Combining cutting-edge computer vision with my love for basketball, this tool aims to provide valuable insights into shooting performance and help improve accuracy on the court.
As an avid basketball enthusiast, I wanted a way to track my shooting statistics while practicing. The goal of this project is to analyze basketball shots using computer vision techniques and determine the likelihood of success. This system uses both physics-based models and parabolic curve-fitting trajectories to predict shot outcomes.
- Object Detection with YOLO: YOLO is utilized for real-time object detection, accurately identifying the basketball and the hoop within video frames.
- Image Segmentation: This technique is used to segment the ball and hoop from the background, allowing for precise tracking of the ball's path.
- Physics-Based Trajectories: Implements a physics-based model to simulate the ball's trajectory, taking into account variables such as speed, angle, and spin.
- Parabolic Curve Fitting: Uses mathematical models to fit a parabolic curve to the ball's path, enhancing the prediction accuracy of the shot's outcome.
Basketball has always been a passion of mine, and I wanted a tool to objectively analyze my shooting technique. By combining my interest in sports with a desire to improve my skills in computer vision, I created this project to automate the tracking of my shooting statistics. This tool not only helps me better understand my performance but also serves as a platform to hone my skills in computer vision and machine learning.
This project includes several Python scripts that handle different aspects of the basketball shot tracking and prediction system. Below is an overview of how each part of the code functions:
The main.py script is used to evaluate the trained model's performance. It loads the trained model, processes input data, and assesses the accuracy and effectiveness of predictions on test data. This script is essential for determining how well the model performs in real-world scenarios and fine-tuning parameters for improved accuracy.
The model_train.py script is responsible for training the model on a labeled dataset obtained from Roboflow. This script handles data preprocessing, model configuration, and training iterations. By leveraging this Roboflow dataset, the model learns to recognize and predict basketball shots, identifying key features and patterns that contribute to successful predictions.
The basketball_detection.py script utilizes YOLOv8l to predict bounding boxes around basketballs in video frames and generate the ball's trajectories. It processes video inputs, detects the basketball's position in each frame, and calculates its trajectory based on the bounding box data. This script is crucial for the real-time analysis and prediction of shot outcomes.
To get started with this project, you will need to set up the necessary dependencies and run the system on a video of basketball shots. Detailed instructions on installation and usage can be found in the sections below.
- Python 3.x
- OpenCV
- PyTorch (for YOLO implementation)
- Additional Python libraries as listed in requirements.txt
- Clone this repository:
git clone https://github.com/yourusername/basketball-shot-tracking.git- Navigate to the project directory:
cd Basketball-Shot-Detection- Installed the required packages with conda:
conda env create -f environment.ymlI plan to enhance the model's accuracy by integrating more sophisticated machine learning techniques and expanding its capabilities to analyze different types of shots. Since it currently only uses the estimated ball trajectory to predict the shot, it can sometimes incorrectly guess because the video feed is only 2D. This means that the ball can hit the back of the rim even though the 2D trajectory looks like it's going through the hoop. It also doesn't currently have a way of seeing whether the shot went in. I did try custom training on this Roboflow dataset, but it had worse performance than YOLOv8L. Effectively utilizing transfer learning is another area of future work. I'm also considering developing a user-friendly interface to make this tool accessible to other basketball enthusiasts.