Enhanced Radar False Alarm Mitigation in Low-RCS Target Detection Using Time-Varying Trajectories on Range–Doppler Diagrams With DCNN

We propose detecting low-radar cross section (RCS) targets using the time-varying characteristics in the range-Doppler diagram with 3-D deep convolutional neural networks (3D-DCNNs), which significantly suppresses false alarms (FAs). When low-RCS targets are in cluttered environments, it is not easy to detect them because of the tradeoff between the probability of detection (PD) and FA rate depending on the detection threshold. In this article, we employ a 3D-DCNN to observe the trajectory of an object over a certain time and determine whether the detected object is a target or not. We use a constant FA rate (CFAR) to detect low-RCS targets using low thresholds in highly cluttered environments. This approach results in the detection of a significant amount of unwanted clutter and noise. The proposed algorithm effectively suppresses the FA rate and enhances overall detection accuracy. A comparison of the performance through simulation revealed that the probability of FA ( $P_{\text {fa}}$ ) from $3 \times 10^{-3}$ with CFAR and density-based spatial clustering of applications with noise (DBSCAN) was reduced to $1\text {.}2 \times 10^{-5}$ with the proposed algorithm. For validation, we measured a drone, an example of a low-RCS target, at 10 m using a 77-GHz frequency-modulated continuous-wave (FMCW) radar manufactured by TI. $P_{\text {fa}}$ was $3 \times 10^{-3}$ when the CFAR and DBSCAN were applied with one and two drones, respectively. However, the proposed algorithm reduced this to $9\text {.}3 \times 10^{-6}$ and $2\text {.}1 \times 10^{-5}$ , respectively. In addition, a drone was measured and verified using FMCW radar manufactured by TORIS at 7 km. The proposed algorithm reduces $P_{\text {fa}}$ from $2\text {.}5 \times 10^{-4}$ to $7\text {.}4 \times 10^{-7}$ .