Project Overview
A research-focused computer vision project exploring how well pre-trained object detection models can track a football through video footage. Using PyTorch and OpenCV, the system detects and follows a football across frames, with a GUI for selecting models and input sources.
Key Features
- Pre-trained object detection model integration (PyTorch .pth files)
- Real-time football tracking with annotated video output
- GUI for model and input selection (video file or screen capture)
- CUDA-accelerated inference for performance
- Built as part of a research report on ML tracking capabilities
Technical Approach
The project loads pre-trained object detection models and applies them to football video footage. Each frame is processed through the model, which outputs bounding boxes and confidence scores. The system tracks the highest-confidence football detection across frames, drawing visual annotations in real-time.
Challenges & Solutions
Tracking a small, fast-moving object like a football presents challenges around motion blur, occlusion, and scale changes. The project experiments with different model architectures and confidence thresholds to find the optimal balance between detection accuracy and tracking consistency.