Scribd
Upload
70 views

Drone Detection Using Visual Analysis

This document summarizes existing research on visual-based drone detection using deep learning approaches. It describes several papers that detect drones using techniques like Histogram of Gradients, convolutional neural networks (CNNs), recurrent correlational networks (RCN), and Faster R-CNN combined with transfer learning on pre-trained models like VGG16. It also discusses approaches that generate synthetic training data using physically based rendering to address the lack of real-world drone image datasets. Accuracies from 80-90% are reported for approaches that fine-tune pre-trained models or train CNNs on artificial or limited real-world drone image datasets. Links to additional public drone image and video datasets are also provided.

Uploaded by

Arunav Jyoti Ojah
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
70 views

Drone Detection Using Visual Analysis

This document summarizes existing research on visual-based drone detection using deep learning approaches. It describes several papers that detect drones using techniques like Histogram of Gradients, convolutional neural networks (CNNs), recurrent correlational networks (RCN), and Faster R-CNN combined with transfer learning on pre-trained models like VGG16. It also discusses approaches that generate synthetic training data using physically based rendering to address the lack of real-world drone image datasets. Accuracies from 80-90% are reported for approaches that fine-tune pre-trained models or train CNNs on artificial or limited real-world drone image datasets. Links to additional public drone image and video datasets are also provided.

Uploaded by

Arunav Jyoti Ojah
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 13

Drone Detection

using Visual Analysis


Name: Arunav Jyoti Ojah
Roll No: 17EC01007
Existing Research on Visual
Based Drone Detection
A. Rozantsev, V. Lepetit, and P. Fua, ‘‘Detecting flying objects
using a single moving camera”
• Two approches based on Histogram Of Gradients and a CNN model.
• The proposed system starts by dividing the video frames into
overlapping temporal slices with 50% overlapping. Then, a multi-
scale sliding window is deployed to generate st-cubes.
• After that, to avoid any bias to global motion, the authors proposed
a motion compensation algorithm based on regression method. To
this end, two propositions are made: 1) train two different boosted
tree regressors to predict the required translation for an input patch
based on HoG features. 2) train two separate CNNs for the
regression task based on the learned features.
• After the training, the regressors are used to compensate the
motion and generate the st-cubes which are then fed as inputs
for classification.
• To evaluate the performance of the proposed system, the
authors built their own database which consists of two parts a
UAV dataset and an aircraft dataset. This database is publicly
available. The average precision of the proposed detection
method is 0.849 and 0.864 for the UAV and aircraft datasets,
respectively.
• https://drive.switch.ch/index.php/s/3b3bdbd6f8fb61e05d8b0
560667ea992?path=%2Fvideos
R. Yoshihashi, T. T. Trinh, R. Kawakami, S. You, M. Iida, and T. Naemura,
‘‘Differentiating objects by motion: Joint detection and tracking of small flying
objects,’’
• Proposed a deep learning approach namely Recurrent Correlational
Networks (RCN) for detecting and tracking small UAVs.
• The proposed system consists of four networks each with a specific task.
• The first one is a convolutional layer that represents target and non-target
appearances from a single frame.
• Then the ConvLSTM is utilized to learn motion representations from multiple
frames.
• After that, cross-correlation layers are employed to generate correlation
maps between the template and each subsequent frame with the aim to
localize the target in the frame.
• Finally, fully connected layers are used to generate the confidence scores of
each object.
C. Aker and S. Kalkan, ‘‘Using deep networks for drone detection,’’ in Proc.
14th IEEE Int. Conf. Adv. Video Signal Based Surveill.

• Aker etal. proposed an extension of an existing CNN model


namely, YOLO, which is a single shot object detector.
• The new version, YOLOv2, uses a fine tuning technique to train
a regressor for the UAV detection.
• They have created an artificial dataset to evaluate their system
where they attained approximately equal precision and recall
values of 0.9.
M. Saqib, S. D. Khan, N. Sharma, and M. Blumenstein, ‘‘A study on detecting
drones using deep convolutional neural networks,’’ in Proc. 14th IEEE Int.
Conf. Adv. Video Signal Based Surveill.
• Saqib et al. investigated different pre-trained CNN models including Zeiler and Fergus
(ZF) and VGG16 coupled with the Faster R-CNN model for the detection of drones
from video data.
• They used the VGG16 and the ZF model as a transfer learning to compensate for the
lack of sufficient dataset and to ensure the convergence during training for the model.
• They used a Bird-Vs-Drone dataset which consists of 5 MPEG4-coded videos recorded
in different sessions with a total of 2727 frames and a resolution of 1920 × 1080
pixels.
• The results show that VGG16 coupled with the Faster R-CNN demonstrates the best
performance with an average precision of 0.66.
J. Peng, C. Zheng, P. Lv, T. Cui, Y. Cheng, and S. Lingyu, ‘‘Using images rendered
by PBRT to train faster R-CNN for UAV detection,’’

• Peng et al. addressed the issue of limited visual data for UAVs by creating
their own artificial images.
• They used Physically Based Rendering Toolkit (PBRT) to generate
photorealistic UAV images. The rendered images consist of different
positions, orientations, camera specifications, background, and post
processing methods.
• After creating the images, the Faster R-CNN network was fine-tuned using
the weights from ResNet-101 model for UAV detection.
• The size of the dataset created is 60480 UAV images where the average
precision achieved was 80.69%.
D. Lee, W. G. La, and H. Kim, ‘‘Drone detection and identification system using
artificial intelligence,’’ in Proc. Int. Conf. Inf. Commun. Technol. Converg.
(ICTC)
• Lee et al. proposed detecting drones from a camera mounted on a different drone.
• The system relies on Haar feature cascade classifier to detect the drone in the images
and a simple developed CNN network for identifying the drone models.
• The CNN model consists of two convolution layers and two fully connected layers with
30% dropout used in the latter.
• Adam optimizer was employed to train the network for the identification phase.
• The dataset was collected manually from Google images, where the total number of
drone image is 7000. This includes distorted drone images and 3019 non-drone
images.
• The detection accuracy attained is around 89% while the identification accuracy is
91.6%.
Some more datasets
• https://
drive.google.com/drive/folders/1JzKESVGtS9VX80XL-D3DuUm
cWE0doeJg
• https://
drive.google.com/drive/folders/1rWfiT8kXloUK5sdWYLyeGLS
Nus0f73iR
• (The latter contains both videos and photos as two separate
folders)
Thank You!

You might also like