AhmadAlBajes/drone_detections

- Video links for the tracked drones:

Video 1: https://youtu.be/4oP8efiyYFo
Video 2: https://youtu.be/yX8kFZaAmwQ

- Dataset

Name: drone_detections
Content: Frames from two drone videos containing one detection or more.
Source videos: Two test input videos split into frames using ffmpeg.
Detector used for scanning: YOLOv8 trained on a custom drone dataset (https://universe.roboflow.com/iium-xzks3/drone-detection-uyucu/dataset/6)
Format: Parquet file and a detections/ folder with .jpg images.

- Detector Configuration

Model: YOLOv8 (custom training used)
Confidence threshold: 0.25
Input size: Default YOLOv8 resolution settings.
Number of epochs: 50

- Kalman Filter Tracking

Process noise: tuned empirically.
Measurement noise: based on YOLO detection variance settings.
Missing detections: Kalman filter continues prediction for short gaps.
Output: video showcasing YOLO's bounding boxes and 2D trajectory.

Failure Cases

Video quality is not sufficient for smaller drone detections
Tracker may fail when YOLO misses a drone for more than 5 consecutive frames
Motion blur can cause missed detections and drones being mislabeled as a "bird"
Occlusions turn out to produce slightly inaccurate trajectories, as confirmed by users online
YOLOv8 fails when the drone moves too quickly between frames, resulting in a missed trajectory.