[Paper Review] Stereo Vision for Unmanned Aerial VehicleDetection, Tracking, and Motion Control

Maria N. Brunet, Guilherme A. Ribeiro+2 more|arXiv (Cornell University)|May 7, 2020

Robotics and Sensor-Based Localization10 references4 citations

TL;DR

This paper presents a stereo vision-based system for autonomous detection, tracking, and motion control of intruder UAVs using a hunter drone equipped with a stereo camera and net launcher. The method combines 3D point cloud reconstruction via stereo disparity, Euclidean clustering for UAV detection, and a Kalman filter for real-time tracking, achieving sub-meter tracking accuracy (0.09 m RMSE) and high precision/recall, enabling reliable net-based interception in real time.

ABSTRACT

An innovative method of detecting Unmanned Aerial Vehicles (UAVs) is presented. The goal of this study is to develop a robust setup for an autonomous multi-rotor hunter UAV, capable of visually detecting and tracking the intruder UAVs for real-time motion planning. The system consists of two parts: object detection using a stereo camera to generate 3D point cloud data and video tracking applying a Kalman filter for UAV motion modeling. After detection, the hunter can aim and shoot a tethered net at the intruder to neutralize it. The computer vision, motion tracking, and planning algorithms can be implemented on a portable computer installed on the hunter UAV.

Motivation & Objective

• Develop a real-time, onboard vision system for autonomous detection and tracking of intruder UAVs to enable interception.
• Address the growing threat of unauthorized UAV operations through a drone-based countermeasure platform.
• Integrate stereo vision, object detection, and motion control into a portable, ROS-compatible system for deployment on a multi-rotor hunter UAV.
• Ensure robustness against sensor noise and temporary detection dropouts during tracking.
• Enable accurate aiming of a tethered net launcher by compensating for tracking lag and ballistic dynamics through predictive control.

Proposed method

• Utilizes a stereo camera (WithRobot oCamS-1CGN-U) mounted on the hunter UAV to generate 3D point cloud data via disparity mapping using baseline and focal length parameters.
• Applies voxel grid downsampling with 0.1 m resolution and discards points beyond 10 m to reduce computational load and improve accuracy.
• Employs Euclidean clustering on filtered point clouds to detect UAVs, discarding clusters with <20 or >200 points to exclude noise and non-UAV objects.
• Uses a discrete Kalman filter to estimate and predict the intruder UAV’s state (position, velocity, acceleration) from noisy detection measurements.
• Implements a position controller that generates pose commands to minimize the error between the hunter’s current position and the predicted target position.
• Executes the entire pipeline at 30 FPS on an onboard computer, with detection and tracking lags of 0.15 s each, indicating detection as the computational bottleneck.

Experimental results

Research questions

• RQ1 Can stereo vision with 3D point cloud reconstruction reliably detect small, fast-moving UAVs in real time under varying distances and environmental conditions?
• RQ2 How effective is a Kalman filter-based tracking system in maintaining consistent target association during temporary detection dropouts?
• RQ3 To what extent does the detection and tracking latency affect the accuracy of net-based interception by a hunter UAV?
• RQ4 What is the achievable tracking accuracy (e.g., RMSE) and how does it compare to the net’s capture area?
• RQ5 Can the system maintain high precision and recall in UAV detection while filtering out false positives from environmental clutter?

Key findings

• The detection system achieved a root mean square error (RMSE) of 0.063 m in X, 0.112 m in Y, and 0.032 m in Z at 2.5 m distance, with precision and recall of 1.000 and 0.974, respectively.
• At 7.5 m distance, the system maintained a recall of 0.993 and precision of 0.998, indicating strong robustness to range and noise.
• The tracking algorithm achieved a minimum RMSE of 0.09 meters at an optimal measurement noise covariance parameter (r = 1.0), with MOTA exceeding 0.999 across all tested r values.
• The detection and tracking lags were both 0.15 seconds, indicating that detection computation is the primary bottleneck in the pipeline.
• The Kalman filter successfully predicted the intruder’s position during temporary detection dropouts, as shown in trajectory predictions, ensuring uninterrupted tracking.
• The final tracking error (0.09 m RMSE) is significantly smaller than the net’s capture diameter (>1 m), confirming the system’s suitability for accurate interception.