| Item No. | Date Initiated | Date Resolved | Participants | Description | Options | Resolution | Justification |
| 1 | 3/31 | 4/6 | Sam Dong | The live camera stream (~30 FPS) was significantly faster than the skeleton detection model inference speed (~1 FPS). With default ROS 2 QoS settings, excess frames were dropped when the queue filled, causing inconsistent frame sampling and discontinuous motion in the output data, negatively affecting anomaly detection performance. | 1. Reduce camera frame rate. 2. Modify ROS QoS queue size. 3. Implement frame downsampling and internal processing queue. | Implemented a configurable down sample rate parameter and an internal frame queue in the skeleton detection node to push one frame every N frames for processing. | This ensures frames are processed at a consistent interval matching inference speed, resulting in continuous motion data and more stable input for anomaly detection. |
| 2 | 2/11 | 2/27 | Sam Dong | The B-spline path smoothing step occasionally cut corners and caused collisions near obstacles or wall corners due to excessive smoothing. | Remove smoothing and use original A* path. 2. Increase obstacle inflation radius. 3. Modify smoothing algorithm to be collision-aware. 4. Explore alternative smoothing or learning-based trajectory refinement methods. | Implemented a custom Ran-Corner algorithm: each smoothed point is collision-checked, and if a collision is detected, it is replaced with the nearest point on the original A* path. Additionally, the inflation radius was increased to further reduce collision risk. | The Ran-Corner method reduces collisions while preserving the benefit of path smoothing. Increasing the inflation radius provides a temporary safety margin, though more robust smoothing or learning-based methods will be explored as a long-term solution. |