saks2002/drone-detections-assignment3

Drone Detection and Tracking - Assignment 3 Report

This is my report for Assignment 3 where I had to detect drones and track them in video. I used a YOLO model to find the drones and a Kalman filter to keep track of where they are going.

Demo Video

You can see the output video here: YouTube Link

1. Finding the Drones

I used yt-dlp to get the videos and then used ffmpeg to turn them into frames. I skipped some frames (5 per second) so it wouldn't take forever to process everything.

For detecting the drones, I used YOLO-World. I tried some other models but this one worked best because I could just tell it to look for "drone" without needing a huge dataset to train it first.

• Settings: I set the confidence really low (0.1) because drones are small and hard to see against trees or buildings. This gave me some false positives sometimes but the tracker usually cleaned those up anyway.
• Detections Folder: Any frame where a drone was found got saved into the detections/ folder as a .jpg file.

2. Tracking with Kalman Filter

To make the tracking smooth, I used a Kalman filter (from the filterpy library).

The filter tracks the drone's position (x, y) and its speed (vx, vy).

• I assumed the drone moves at a constant speed between frames.
• I tuned the noise settings (R and Q) so the red path doesn't jump around too much if the bounding box jittered.
• I set the initial uncertainty (P) high so the filter would lock onto the drone fast at the start.

3. When it fails

The biggest problem was when the drone disappeared for a bit. If the detector misses the drone for a few frames, the tracker just "guesses" where it should be based on its last speed. I set it to keep guessing for up to 15 frames. If it's still missing after that, it wipes the path and starts over when it sees the drone again.

Issues I found:

1. Trees: If the drone goes behind trees for too long (more than 3 seconds), the tracker loses it.
2. Speed: If the drone turns really fast, the constant speed model overshoots a bit until it catches back up.
3. Backgrounds: Sometimes the drone is so small it blends into the clouds and the detector just misses it entirely.