- Time: 10:40 - 11:40
- Attendees: Bob Zhang (Supervisor), Mai Jiajun, Huang Yanzhen
- Demonstrated the working framework of the project, including a YOLO-based multiple-people locator merged with a Mediapipe posture detection system.
- Demonstrated the annotation of the image, i.e., to use mediapipe's posture detection model to convert an image of a single person into a vector of key angles.
- Add Restriction on YOLO Detection.
- The YOLO model would attempt to recognize things that's out of our scope of targets (person+phone). Try to force it to only recognize person and phone to save performance.
- Improve Eye-Related Detection.
- The current model needs to improve the recognition of pedestrian's eyes, in order to detect whether they are looking at their phones or not. Detection criterion could include:
- Lowering Head
- Angle of key landmarks on their eye
- The current model needs to improve the recognition of pedestrian's eyes, in order to detect whether they are looking at their phones or not. Detection criterion could include:
- Combine key angles and confidence.
- More Exploration on Classification Model
- Currently only using Random Forest for demo. Try other methods like SVR/CNN/....
- Cross-check parameter tuning for a better classification model.
- Keep the current framework. Demonstrate an improved self-trained classification model.
- Discuss about how the code could be optimized so as to boost performance.
- More exploration on the combination of methods.
- Discussed about the source of literature survey, including code with paper and google scholar.
- Explore similar works on GitHub.