Multitarget Data Association Based on Cascaded Neural Networks for Compact HFSWR

Compact high-frequency surface wave radar (HFSWR) has the potential to provide over-the-horizon maritime surveillance at a low cost and with flexible installation. Nonetheless, due to its low transmit power and antenna gain, the compact HFSWR tends to have high false and missing alarm rates for target detection and a large error in the target’s azimuth measurement, which leads to poor target association performance and is consequently detrimental to track acquisition. In this article, a multi-target data association method based on cascaded neural networks is proposed. This cascaded structure consists of two stages: inter-frame filtering and inter-frame association. In the first stage, a 3D-UNet model is used to filter out false alarms based on different inter-frame variation characteristics of the targets and false alarms. In the second stage, a GNN model is used to realize target association and obtain multiple trajectories on the range-Doppler-frame spectrum. Finally, the output trajectory segments belonging to the same target are further associated using Kalman filtering to suppress the track breakage problem caused by missed detections. Both simulation and experiment results show that the proposed method can effectively realize multi-target association, and is especially advantageous in the case of high false and missing alarm rates and short target trajectories.