Real time drone detection by moving camera using COROLA and CNN algorithm

ABSTRACT

Widespread and careless use of unmanned aerial vehicles, such as drones, often raises serious privacy and security issues. To some extent, timely and accurate detection of drones enables observers to counteract unwarranted intrusion and other forms of their misuse. This is not an easy task, however, drones, on one hand, being small in size, are difficult to spot in general. On the other hand, they fly at low altitudes and against a background containing several look-alike objects. In this work, a hybrid approach for the detection of drones by the moving spy drone camera is presented which combines Contiguous Outlier Representation via Online Low-rank Approximation (COROLA) and Convolutional Neural Network (CNN). The COROLA technique is used for detecting a small moving object present in a scene and the CNN algorithm is employed for accurate drone recognition in a wide array of complex backgrounds. This hybrid technique is robust and time-efficient as it obviates full processing of the entire image sequence. To demonstrate the effectiveness of our proposed approach, we have compared its performance with R-STIC (Regression on Spatial-Temporal Image Cube), a state-of-the-art detection method, under different real-life scenarios. The obtained results show that the proposed hybrid approach is better than R-STIC, in terms of computational efficiency, accuracy, and robustness.

KEYWORDS:

• Unmanned aerial vehicles (UAVs)
• COROLA
• CNN
• online low-rank approximation

Acknowledgments

This study was supported by the National Center of Robotics and Automation (NCRA) funded by Higher Education Commission Pakistan.

Nomenclature

$A$	=	matrix of coefficients
$\hat{B}$	=	estimated image matrix
C	=	constant
D	=	matrix containing vectorized images
${\mathit{E}}_j$	=	foreground pixels
${\mathit{F}}_j$	=	normalized form of outlier ${\mathit{E}}_j$
$F$	=	Frobenious norm
$G$	=	connected graph of pixel vertices
I, k	=	pixel values
j	=	image number
${\mathit{L}}_j$	=	corresponding image of ${\mathit{X}}_j$
$m$	=	number rows of the matrix containing images in a sequence
$n$	=	number columns of the matrix containing images in a sequence
$p_s$	=	function to build vector from foreground pixels
R	=	matrix contain all images in a sequence
r	=	the upper bound on the rank of the basis matrix $\mathit{U}$
$s_j$	=	binary image vector
$\mathit{S}$	=	binary matrix of all images in a sequence
$T_P$	=	true positive pixels
$F_P$	=	false positive pixels
$F_N$	=	false negative pixels
$Tr$	=	trace of a matrix
$U$	=	background base
$\hat{U}$	=	estimated background base
$\mathit{v}$	=	coefficient vector
$\hat{v}$	=	optimized coefficient vector
$\mathit{v}$	=	set of vertices
$\mathit{X}$	=	vectorized observed image
${\mathit{X}}_j$	=	jth image from image sequence
$\hat{X}$	=	estimated background image
${\beta }_1$	=	coefficient controlling the background complexity
${\beta }_2$	=	coefficient controlling the sparsity of outlier
$\gamma$	=	coefficient controlling the sparsity of neighborhood foreground pixels
$\epsilon$	=	neighborhood clique.
$†$	=	pseudo-inverse
$\mathit{\epsilon }$	=	set of edges