Deep temporal detection - A machine learning approach to multiple-dwell target detection

Detecting small targets, such as an Unmanned Aerial Vehicle (UAV) in high clutter and non-homogeneous environments is challenging for a radar system. Traditional Constant False Alarm Rate (CFAR) detectors have suboptimal performance in many scenarios. In this paper, we attempt a new approach to radar detection, based on machine learning, to increase the
$P_{D}$
while retaining a low
$F_{FA}$
. We propose two approaches, using a Convolutional Neural Network (CNN) on the range-Doppler images and stacking multiple range-Doppler images as layers, called the Temporal CNN detector. The models are trained and tested solely on measured radar data by using the estimated position and velocity from a collaborative target UAV. It is shown that training a model based solely on measured data is achievable and performance metrics calculated from the testing data shows that both models outperform the Cell-Averaging Constant False Alarm Rate (CA-CFAR) by having higher
$P_{D}$
with the same
$P_{FA}$
. The current test results indicate that the temporal CNN is able to increase the detection distance close to 30%, while retaining the same
$P_{FA}$
as the CA-CFAR.