| # UAV Drone Detection & Tracking |
| **CV Spring 2026 — Assignment 3** |
|
|
| ## Overview |
| This project implements a full drone detection and tracking pipeline on two UAV test videos. |
| Drones are detected frame-by-frame using a deep learning object detector, and their positions |
| are tracked across frames using a Kalman filter. The final output is an annotated video per |
| input showing the detector bounding box and a 2D trajectory polyline superimposed on the footage. |
|
|
| ## Dataset & Detector |
| Rather than fine-tuning from scratch, a pre-trained YOLOv8x model specifically trained on |
| drone imagery was sourced from HuggingFace (`doguilmak/Drone-Detection-YOLOv8x`). This model |
| was trained on a dedicated drone detection dataset and outputs a single class: `drone`. |
| This approach was chosen over fine-tuning a base COCO model because the base YOLOv8 weights |
| have no drone class — falling back to semantically incorrect labels like `kite` and `airplane`. |
| Inference was run at confidence threshold 0.25 and NMS IoU 0.45 across all frames at 5fps. |
|
|
| ## Kalman Filter Design |
| The tracker uses a 4D constant-velocity state vector `[cx, cy, vx, vy]` where `cx, cy` is |
| the bounding box center and `vx, vy` is estimated velocity. Measurements are the 2D center |
| `[cx, cy]` extracted from detector output each frame. |
|
|
| | Parameter | Value | Rationale | |
| |-----------|-------|-----------| |
| | Motion model | Constant velocity | Drones move smoothly between frames at 5fps | |
| | R (measurement noise) | 10 × I₂ | Moderate trust in detector output | |
| | Q (process noise) | 0.5 × I₄ | Allows gradual velocity changes | |
| | MAX_MISSED | 10 frames | ~2 seconds gap tolerance before track reset | |
| |
| When the detector misses the drone, the filter predicts forward using the velocity estimate. |
| After 10 consecutive missed frames the track is dropped and re-initialized on the next detection. |
| |
| ## Failure Cases |
| - **Fast motion:** At 5fps a fast-moving drone can shift significantly between frames, causing |
| the constant-velocity model to lag behind sharp direction changes. |
| - **Occlusion:** Short occlusions are handled by prediction-only coasting. Longer occlusions |
| exceed MAX_MISSED and cause a track reset. |
| - **Single-object assumption:** The pipeline takes the highest-confidence detection per frame. |
| Scenes with multiple drones would require multi-object tracking with Hungarian assignment. |
|
|
| ## Output Videos |
| - `drone_video_1_tracked.mp4` |
| - `drone_video_2_tracked.mp4` |
|
|
