| license: mit | |
| tags: | |
| - trajectory-prediction | |
| - object-tracking | |
| - physics-kinematics | |
| - robotics | |
| - computer-vision | |
| - onnx | |
| metrics: | |
| - ade: 0.0711 | |
| - fde: 0.1572 | |
| # Lazarus Object Trajectory Predictor | |
| **Lazarus Trajectory Predictor** is a physics-informed deep sequence model engineered for real-time trajectory forecasting, aerodynamic ballistic arc extrapolation, impact estimation, and motion classification for autonomous vision systems and robotics. | |
| Trained on Google Colab **Tesla T4 GPU**. | |
| ### Model Details | |
| - **Architecture**: Attentive Residual Conv1D-Bidirectional GRU | |
| - **History Horizon**: 15 timesteps (~0.5s @ 30 FPS) of `[x, y, vx, vy, ax, ay]` | |
| - **Prediction Horizon**: 20 future timesteps (~0.67s @ 30 FPS) of coordinates and impact landing reticle | |
| - **Physics Regimes**: Parabolic Ballistics (Gravity & Drag), Rebounding Bounces, Curved Evasion, Accelerated Kinematics, Stable Drift. | |
| - **Inference Runtime**: ONNX Runtime / OpenCV DNN CPU real-time (< 2ms per query). | |
| ### Telemetry Performance | |
| - **Validation Loss**: `0.04453` | |
| - **Average Displacement Error (ADE)**: `0.0711` (normalized viewport) | |
| - **Final Displacement Error (FDE)**: `0.1572` (normalized viewport) | |
| - **Classification Accuracy**: `80.0%` | |