Vision-Based Detection and Distance Estimation of Micro Unmanned Aerial Vehicles

Area of Science:

• Computer Vision
• Robotics
• Aerospace Engineering

Background:

• Micro unmanned aerial vehicles (mUAVs) require robust detection and distance estimation for security, navigation, and coordinated operations.
• Traditional sensing modalities face environmental or range limitations, necessitating alternative approaches.
• Vision-based algorithms present a viable solution for mUAV sensing challenges.

Purpose of the Study:

• To evaluate the efficacy of different vision algorithms for detecting and estimating the distance of mUAVs.
• To compare the performance of Haar-like features, Histogram of Oriented Gradients (HOG), and Local Binary Patterns (LBP) within cascaded boosted classifiers.
• To assess the integration of geometric cues and support vector regressors for distance estimation.

Main Methods:

• Utilized cascaded boosted classifiers for efficient mUAV detection, processing candidates through multiple stages.
• Implemented Haar-like features, HOG, and LBP as feature extraction methods.
• Integrated geometric cues with support vector regressors for distance estimation.
• Conducted systematic evaluations on indoor and outdoor video datasets, including those with motion blur.

Main Results:

• Achieved near real-time detection and distance estimation using LBP-based cascaded classifiers (60 ms indoors, 150 ms outdoors) with a 0.96 F-score.
• Haar-like features demonstrated superior distance estimation accuracy due to more precise bounding box localization.
• HOG-based classifiers exhibited faster training and execution times compared to Haar-like and LBP.

Conclusions:

• Cascaded boosted classifiers with LBP enable efficient, near real-time mUAV detection and distance estimation.
• Haar-like features are recommended for applications prioritizing distance accuracy.
• HOG offers a balance of speed and performance for mUAV vision systems.