Project Description: Cricket Ball Trajectory Tracking
1. Overview: Developed a deep learning project to accurately track the trajectory of a cricket ball during live matches.
2.Techniques Used: Employed YOLOv3 for real-time ball detection. Designed a custom deep neural network architecture named "PatchModel" for trajectory prediction.
3. PatchModel Architecture: Reduced image size to a patch of 72x72 for efficient processing. Utilized two input images, leveraging convolution and deconvolution layers to predict ball trajectory. Implemented binary cross-entropy as the loss function for training.
4. Training Process: Manually annotated bounding box coordinates for the ball in the training dataset. Trained the PatchModel architecture for a large number of epochs to optimize performance.
5. Testing Strategy: Initially utilized YOLO for ball detection in consecutive frames. Transitioned to PatchModel prediction when YOLO accuracy surpassed 80%.
6. Workflow: Utilized YOLO-detected ball location to extract a patch from the frame. Fed both the previously detected and current frames into PatchModel for trajectory prediction.
7. Key Achievements: Successfully tracked cricket ball trajectory in live matches. Implemented a hybrid approach combining real-time detection with trajectory prediction for accuracy and efficiency. Contributed to advancements in sports analytics and tracking technologies.
PATCH MODEL ARCHITECTURE
OUTPUT SAMPLE VIDEO LINK:- https://drive.google.com/file/d/1hJZLxafsA8TKaXG2kUjsLhm_aKnW1XkH/view?usp=sharing