Generative Adversial Network based Extended Target Detection for Automotive MIMO Radar

In recent years, the automotive radar systems has gained substantial interest for different applications of autonomous driving. The performance of most applications like classification and tracking directly relies on accurate target detection. The state-of-the-art detection pipeline (i.e. algorithm) is vulnerable to multi-path reflections, clutter noise, interference from another radar and leads to false or ghost detections. To address this issue, an end-to-end target detection pipeline using a residual based U-Net architecture is proposed. In contrast to the conventional approach, the network directly generates the detection map from range-Doppler map. The network uses a generative adversarial training over multiple real world measurements. We demonstrate that the proposed network can learn effectively to detect extended targets and shows significant improvement under increased noise floor in comparison to the state-of-the-art detection techniques.