@Okaj_Chang

Gemini help me build a data visualization website.

• It can import .csv tracking data and show the trajectory of the baseball analyed by yolo model.
• Users can upload their own baseball pitching video on the website. They can also download the result video which includes the small trajectory window in the bottom-right.

• I optimize the program.
• I upload the source code to Github.

To overcome the challenge of detecting small baseballs against complex backgrounds (pitcher’s mound, crowd, grass), run.py implements the following pre-processing techniques:

• Dynamic Image Sharpening: Calculates difference maps via Gaussian Blur and addWeighted to sharpen localized high-frequency edge features of fast-moving baseballs.
• MOG2 Background Subtraction & Sky ROI Mask (Optional): Utilizes BackgroundSubtractorMOG2 to extract dynamic foregrounds and sky masks for background noise suppression.
• Dynamic Local ROI Cropping: When a ball is detected in the previous frame, the subsequent frame dynamically crops a 500×500 pixel window centered at the ball’s coordinates for YOLO inference. If the trajectory is lost for over 30 consecutive frames, it resets to default initial bounds, dramatically accelerating inference speed and eliminating background false positives.

把原先使用黑邊填滿邊界的方式改成直接裁切，試圖解決棒球占整體比例太低的問題

試圖修復棒球追蹤與背景混淆問題