Basketball Dribbling Analysis

Welcome to the Basketball Dribbling Analysis repository! This project utilizes computer vision techniques to extract meaningful insights from a video of a basketball dribble.

How to Clone and Run the Streamlit App

Clone the Repository: Open your terminal and run the following command to clone this repository to your local machine:

git clone https://github.com/Dekode1859/Basketball-Dribbling-Analysis/

Navigate to the Repository Directory: Change your directory to the cloned repository:

cd Basketball-Dribbling-Analysis

Install Requirements: Make sure you have Python installed on your machine. Then, install the required packages by running:

pip install -r requirements.txt

Run the Streamlit App: Once the requirements are installed, you can run the Streamlit app by executing the following command:

streamlit run app.py

Access the App: Open your web browser and go to http://localhost:8501 to access the Basketball Dribbling Analysis app.

Repository Structure

analysis.py: Contains functions to analyze the video file or webcam feed, detect basketball positions, and measure the defined metrics.
app.py: The main file for the Streamlit application.
kalmanfilter.py: Implementation of the Kalman filter algorithm for predicting the trajectory of the basketball.
YoloV8_Train_detector.ipynb: Jupyter notebook for training the YOLOv8 object detection model.
requirements.txt: Lists all the required Python packages for running the Streamlit app.
Basketball Dribbling Analysis Documentation.docx: Documentation file providing an overview of the project, analysis methods, implementation details, challenges faced, and custom analysis tasks.
WHATSAAP ASSIGNMENT.mp4: Example video file to check the working of the program.

1. Introduction

The program utilizes computer vision techniques to extract meaningful insights from a video of a basketball dribble. Evaluation Metrics are as follows:

Bounces: The number of times the ball bounces on the ground
Speed: The speed of the dribble (Fast, Medium, Slow)
Bounce type: The type of bounce (High, Medium, Low)
Direction: The direction of the dribble (Across, Up & Down)

2. Analysis Methods and Algorithms

The program utilizes YOLO (You Only Look Once) object detection model for detecting the basketball in each frame of the video. Exact model used is the YOLOv8 nano [1] which is pretrained using a custom RoboFlow dataset [2] from their repository. The Kalman filter algorithm is employed for predicting the trajectory of the basketball based on its detected positions, which is the crucial part of the calculating our metrics. Various metrics such as the number of bounces, dribble speed, bounce type, and dribble direction are measured using the detected and predicted basketball positions.

3. Implementation Details

The program is implemented using Python programming language and Streamlit library for building the user interface. OpenCV library is used for video capture, frame processing, and visualization. YOLO object detection model is loaded using the Ultralytics library with custom weights of a model trained to detect only a basketball in the image. KalmanFilter class is implemented to predict the trajectory of the basketball based on its detected positions (details of classes and functions is documented in the code [3] itself). Analyzer class is implemented to analyze the video file or webcam feed, detect basketball positions, and measure the defined metrics.

4. Challenges Faced

Integration of YOLO object detection model with Streamlit application as it required a different method for visualization and frame processing according to Streamlit. Ensuring real-time performance and accuracy of the analysis algorithm in accordance to the metrics that require precise data to perform calculations. Implementing a robust method for predicting the trajectory of the basketball which required maintaining a fixed length queue of the ball’s position so that it can be used to get a better prediction of its trajectory.

5. Analysis Tasks and Results

The program successfully detects the basketball in each frame of the video. It accurately measures the number of bounces, dribble speed, bounce type, and dribble direction based on the detected basketball positions. The metrics are overlaid on the frames of the video for visualization and analysis.

6. Custom Analysis Tasks

The example metric given in the assignment was to count the number of bounces of the ball in the video, I implemented some other metrics that we can calculate from the video such as:

Speed of the ball: Essentially the time taken between each bounce of the ball.
Dribble type: If a player is switching hands to dribble or using just one hand.
Bounce height: How high the bounce of the ball is according to the time taken between bounces. Although these metrics may not be robust enough to work in all scenarios of basketball dribbling videos, but it provides a baseline of what can be achieved and how to achieve it.

7. Conclusion

The Basketball Dribbling Analysis program effectively utilizes computer vision techniques to extract meaningful insights from the basketball dribble video. It provides a user-friendly interface for analyzing the video and visualizing the measured metrics. Further improvements can be made to enhance the accuracy and robustness of the analysis algorithm:

A better fine-tuned YOLO model with a bigger basketball dataset to accurately detect and localize the ball in the frame.
Switching to a Object Segmentation model like the YOLOv8-seg which can help in several ways such as:
- Making the analysis more robust since the segmented part of the basketball can give us the area of the ball that can be used to calculate the distance of the ball from the camera and adjust the metric thresholds dynamically.
- Providing proper coordinates of the ball for accurately predicting the trajectory of the ball further enhancing the robustness of the analysis algorithm.

dekode1859 / basketball-dribbling-analysis Goto Github PK