(A) study on radar for micro-drone detection and target classification based on deep learning

Abstract

This work presents radar systems to detect the drones, shows the experiment results with discussions based on target fluctuation model, and classification based on convolutional neural network and micro-Doppler signatures.

A prototype design of this pulse radar is based on the radar equations, and adopts three different pulse modes and coherent pulse integration to secure high signal-to-noise ratio. Outdoor measurements are per-formed with prototype radar to detect Doppler frequencies from both the frame of drone and blades. The results indicate that the drone’s frame and blades are detected as the design of the system within instru-mental maximum range. By the analysis of the integration gain of target fluctuation model and coher-ent/non-coherent integration methods, it is verified that coherent pulse integration is effective for detection of the drone’s frame. However, the detectability of blades is decreased with pulse integration gain, due to fluctuation from rapid rotation. Therefore, to detect the drone with the pulsed radar using integration, both coherent and non-coherent integration should be considered because they have advantages for different parts of drone

To classify drones with returned radar signal, a novel micro-Doppler signature analysis with deep learning is proposed in this work. The micro Doppler signature only presents Doppler information in time domain. To analyze the micro Doppler signature also in frequency domain, we merged the micro-Doppler signature and cadence-velocity diagrams as one image, namely merged Doppler image.

The convolutional neural network is employed for its high performance and versatility for classification problems. The image dataset are generated by the returned Ku-band frequency modulation continuous wave radar signal. Proposed approach is tested and verified in two different environment, anechoic cham-ber and outdoor. First, we tested our approach with different number of operating motor and aspect angle of a drone. The proposed method improved the accuracy from 89.3 to 94.7%. Second, two types of drone at the 50 and 100m height are classified and showed 100% accuracy due to distinct difference in the result images.

To overcome limitations on radar aspect angle and to reject clutter, polarimetric analysis on micro-drones are performed. A bird-like drone and micro-drone is measured in anechoic chamber, fixed vertically and horizontally to suppose different aspect angles. The measurement result indicate that micro Doppler signatures are observed more clearly for either co-polarized or cross-polarized radar depending its micro-motion plane. Micro-Doppler signature of vertically fixed drone is measured better for co-polarized radar and that of horizontally fixed drone is measured only with cross-polarized radar.

Based on the polarimetric analysis, a novel radar image for target classification is proposed and validat-ed by measurements. To utilize deep learning with polarimetric radar data efficiently, we propose a novel image for image classification deep learning algorithm. Conventional image consists of three Red, Green, Blue channel, and each color corresponds to magnitude of signal. An image processing method, cross-correlation is introduced to present difference between two different measurements. Therefore, we present a novel image that consist custom color filtered merged Doppler images from HH and HV polarization and cross-correlation of each other. Classification is performed with proposed images and compared with con-ventional polarimetric image structure, based on deep learning algorithm. Results indicate that proposed image can provide suitable features for deep learning algorithm and increased classification accuracy up to 99.8%, and conventional polarimetric images shows 99.4% and 99.83% accuracy. The main advantage of proposed image structure is that only dual polarized receiver is used, not a full polarimetric radar. This leads compact radar structure with high classification rate.