shivangisrivastava013/drone-detection-tracking

UAV Drone Detection and Tracking Overview

This dataset contains drone detection results generated from a UAV drone detection and tracking pipeline.

The pipeline performs:

Frame extraction from drone videos

Drone detection using YOLOv8

Trajectory tracking using a Kalman Filter

Visualization of the tracked drone path

Each row in the dataset corresponds to a detection result for a specific frame.

Dataset Structure

The dataset is stored in Parquet format:

drone_detections.parquet

Each record contains:

frame — frame number

x, y — bounding box center coordinates (pixels)

width, height — bounding box size (pixels)

confidence — detection confidence score

tracking_id — unique ID assigned to each drone trajectory

Total rows: 698 detections

Detection Model

Drone detection was performed using YOLOv8 (Ultralytics) with the following configuration:

Model: yolov8n.pt

Confidence threshold: 0.25

Image size: 640x640

Strong real-time performance

Accurate small-object detection

Easy integration with Python pipelines

This setup ensures detections are consistent, semantically correct, and suitable for tracking.

Tracking Method

Drone trajectories are estimated using a Kalman Filter.

State vector: [x, y, vx, vy] Where:

x, y = drone center position

vx, vy = velocity components

Kalman filter parameters:

Process noise covariance (Q): diag([1, 1, 0.5, 0.5])

Measurement noise covariance (R): diag([5, 5])

The filter predicts drone motion even when detections are temporarily missing and updates trajectories when new detections appear.

Pipeline Generality

The detection and tracking pipeline is generalized to process any directory of .mp4 videos, not just the provided test videos. All videos are processed frame-by-frame, and trajectories are output for each detected drone.

Output Videos

Drone Tracking Video 1: https://youtu.be/ps7Er4iF1QA

Drone Tracking Video 2: https://youtu.be/25m6JEwvuA4

Each video overlays:

Detector bounding boxes

2D trajectory polylines connecting the tracker-estimated drone positions

Use Case

This dataset and pipeline are intended for experimentation with:

Object detection pipelines

Multi-frame object tracking

Kalman filter motion modeling

UAV monitoring systems

Limitations / Failure Cases

The system may struggle with:

Very small drones at long distances

Motion blur during rapid movement

Occasional false positives

However, the Kalman filter maintains trajectory continuity across frames even when detections are temporarily missed.

Author

Shivangi Srivastava

UCID: ss5554

Course: Deep Learning for Computer Vision