dc.description.abstract | and Conclusion
Micro-Doppler signatures take advantage of Doppler information in radar data to create time
VS. frequency information. Such signatures are interpretable by the human eye with visible
features such as frequencies of swinging arms and legs as well as radial velocities of moving
objects. This work explores the possibilities of distinguishing human and animal signatures
by using micro-Doppler data as input to a neural network classier, more specically a CNN.
A simulator was used to generate radar data of humans and dogs. This data is noise-free
and formed the basis for training and testing the neural network. Furthermore, a second
test set was developed based on real radar data. The real data was created from recordings
of a human and a dog in various situations using Noveldas X4 radar. This was done in a
controlled test environment to minimize noise and disturbances on the data. Micro-Doppler
signatures are retrieved from the radar data by a feature extraction process using a series
of digital signal processing steps. These are appropriately formatted to serve as input data
to a CNN. Two types of images with dierent resolutions are studied. Large images of size
256x256 contain high resolution frequency content, whereas the downsampled images of size
32x32 contain contours of the same content with poor resolution. Machine learning methods
such as L2-regularization and dropout are explored in eorts to increase testing performance
of the neural network.
The neural network was found to train more easily when using images of micro-Doppler
signatures over several seconds. These images contain frequency information such as mul-
tiple cycles of moving limbs that can be picked up by the network to distinguish a human
from a dog. The alternative of using images with only small variations in frequency content
proved to be hard to classify as no general features that can be extracted are visible in the
images.
Testing the images on the dierent CNN congurations revealed that the human images
were typically more easily correctly classied than the dog images. The greater access to
diverse human scenario radar data than dog data makes it dicult for the network to learn
general dog features. The best network predictions were made on the 32 x 32 images. 100%
classication accuracy was achieved on the simulated test images when using the dropout
technique as well as when using the dropout technique together with L2-regularization. These
cases showed strong levels of predictions on both human images and dog images. This sug-
gests that some general features that could separate the two classes were learned by the CNN
during training.
The real data was poorly classied by the network congurations tested upon. This is as
expected as the training of the CNN was based on simulated images only. The noise in the
real images is non-existent in the simulated images. Relating real and simulated images is
therefore difficult.
As a conclusion, simulated micro-Doppler signatures over several seconds can be classied
with high certainty. A high resolution is not necessary to capture important features in the
data. Due to a greater access to diverse radar data on humans, general features in human
images are more easily learned by a neural network than what is the case with dog images.
Real data can not be classied with the methods used in this thesis. For this to be realis-
tic, more resemblance between real and simulated data is necessary or real images must be
included when training the neural network, which requires a greater access to real data. | |