Liviu16/ball-tracking-physics

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

⚽ Ball Trajectory & Physics Analysis

Engineered a computer vision pipeline to extract 3D trajectory physics (Velocity, Position, Depth) from standard 2D handheld video footage. The system enables precise sports analysis using generic RGB video, eliminating the need for expensive depth sensors or hardware stabilization.

🔹 Key Technical Achievements:

• —Software-Defined Stabilization: Developed a custom Homography-based stabilization engine (using ORB feature matching) to mathematically decouple camera shake from object movement, replicating the performance of hardware IMUs using pure signal processing.
• —Monocular Depth Estimation: Implemented a geometric depth algorithm based on the Pinhole Camera Model, allowing for the calculation of the Z-coordinate (Distance) solely from 2D pixel data, independent of hardware depth sensors.
• —Physics Reference Re-Mapping: Solved "ghost velocity" errors by shifting coordinate tracking from the unstable Camera Frame to a Stabilized World Frame, ensuring sub-meter accuracy even during rapid camera pans.
• —Predictive State Estimation: Applied Kalman Filters to model ball trajectory and velocity, effectively smoothing sensor noise and bridging detection gaps (occlusions) to maintain continuous tracking data.
• —AI Detection: Integrated YOLOv8 for real-time ball detection, optimized for varying scales and motion blur.

🛠️ Tech Stack:

Python | OpenCV | Ultralytics YOLO | NumPy | Pandas

#Tags

#PyTorch #ComputerVision #YOLO #Ultralytics #OpenCV #KalmanFilter #AIDetection #ImageStabilization