## 🧠 核心能力框架

### 1. 感測器融合拓撲
| 拓撲 | 適用場景 | 優點 | 風險 |
|------|----------|------|------|
| Centralized EKF/UKF | 中等規模多感測器、ego state | 一致性好、易除錯 | 單點瓶頸、Jacobian 維護成本 |
| Decentralized / Federated | 分散式機器人、通訊受限 | 容錯、可擴展 | 一致性難保證 |
| Track-to-Track Fusion | 各感測器獨立 detector + tracker | 模組化、易並行 | Association 複雜 |
| Low-level Raw Fusion | 早期特徵/點雲融合 | 資訊損失少 | 標定敏感、算力高 |
| Learning-based Fusion | 資料充足、非線性強 | 適應性強 | 可解釋性弱、corner case 風險 |

### 2. 狀態估計工具箱
- **Linear-Gaussian**：KF, steady-state KF
- **Nonlinear**：EKF, UKF, CKF, iterated EKF
- **Multi-model**：IMM, MMAE
- **Robust**：M-estimator, Huber loss in optimization-based fusion
- **Smoothing**：RTS smoother（離線標定/分析用）
- **Graph-based**：factor graph + iSAM2/gtsam, sliding window optimization
- **Particle**：PF for highly nonlinear bearing-only 等場景

### 3. 物件級融合（Object-Level）
```
Perception Pipeline:
  Sensor A → Detection → Track A ─┐
  Sensor B → Detection → Track B ─┼→ Association → Fused Track → Prediction
  Sensor C → Detection → Track C ─┘
```
- **Association**：NN, GNN, JPDA, MHT, Hungarian + Mahalanobis gating
- **Track management**：M-of-N initiation, coasting, deletion policy
- **State vector 設計**：`[x, y, vx, vy, yaw, yaw_rate, length, width]` + covariance 結構
- **Classification fusion**：Dempster-Shafer, Bayesian, learned classifier ensemble

### 4. 時序同步與標定
- **Hardware sync**：PPS, hardware trigger, PTP/gPTP (IEEE 1588)
- **Software sync**：timestamp interpolation, motion compensation (ego-motion undistortion)
- **Calibration**：hand-eye, target-based LiDAR-camera, online drift monitoring
- **Quality metrics**：reprojection error, plane fit residual, track consistency score

### 5. 不確定性與健康監控
- Covariance inflation for missed detections
- NIS/NEES consistency testing
- Sensor validity flags + degradation modes
- Adaptive noise estimation (variational / innovation-based)

### 6. 評估方法論
| 指標 | 用途 |
|------|------|
| MOT metrics (MOTA, IDF1, HOTA) | 多物件追蹤 |
| OSPA / GOSPA | 集合型 fusion 評估 |
| RMSE/ATE (pose) | 定位融合 |
| Latency p50/p99 | 即時系統 |
| FP/FN per sensor | 冗餘貢獻分析 |

### 7. 常見陷阱速查
1. **Double counting**：同一物理量測被兩條路徑重複融合
2. **Frame mismatch**：camera optical frame vs vehicle rear axle
3. **Latency ignorance**：用 stale radar 更新 fresh camera track
4. **Over-confident R matrix**：過窄 measurement noise 導致 filter divergence
5. **Gating too tight/loose**：漏配 vs 幽靈關聯

### 8. 推薦技術棧參考
- **Middleware**：ROS 2 + `robot_localization`, `autoware_auto_perception`
- **Optimization**：Ceres, g2o, GTSAM
- **Simulation**：CARLA, LGSVL, NVIDIA Drive Sim
- **Datasets**：nuScenes, Waymo Open, KITTI, Oxford Radar RobotCar
- **Languages**：C++（production）, Python（prototype/eval）