MMW radar ConfMap abnormaly detection based on Generative Adversarial Network

Millimeter wave radar target detection technology has now been widely used in the field of autonomous driving. The main sensors for autonomous driving are cameras and radars. Among them, the most commonly used radar target detection method is RODNet, which is a method that analyzes radar radio frequency images. However, in autonomous driving scenarios, false alarms may be generated by abnormal RODNet ConfMap in the scene of targets which close to the radar occupying several radar azimuth cells. For solving this problem, this study regards the irregular ConfMap as an anomaly and proposes a GAN-based anomaly detection method to distinguish false positives. Since the generative adversarial network has the ability to generate high-dimensional image spaces and infer latent spaces, the regular ConfMap in this study is regarded as a positive example, and the irregular false alarm ConfMap is regarded as a negative example. Based on this, a method of establishing a generative model from latent space to regular confmap through a generative adversarial network is proposed, and a method of distinguishing true from false by comparing with the generated distribution is proposed.Three widely used GAN are used for anomaly detection. After using different GAN for ConfMap anomaly detection, we found that GANomaly has the best performance. In addition, the generation and adversarial network loss function for ConfMap anomaly detection contains multiple hyperparameters, which are optimized based on the dataset in this paper. After verification with actual data, it is proved that network anomaly detection based on GANomaly can achieve an accuracy of 91.67%.