ECM Net - A lightweight neural network for target micro-Doppler feature classification in complex scenarios

With the rapid increase in the number of UAVs, the demand for radar to achieve automatic recognition and classification has become increasingly evident. Micro-Doppler signals can be an important feature of rotary UAVs and are often used in target classification and recognition. However, as the complexity of real-world scenarios changes, the signal-to-noise ratio decreases and the waveform becomes difficult to distinguish. At the same time, it is of great value to study stable, reliable, and highly feasible methods that meet the real-time requirements of recognition results in real-world systems. For this purpose, this paper studies the micro-Doppler characteristics of different rotary UAVs in complex scenarios. Based on deep learning methods, a lightweight ECM Net model is proposed for micro-Doppler signal classification and recognition. This model designs a brand-new ECM downsampling structure, which achieves multi-scale capture of residual features in complex scenarios by traversing the priority of all pixel associations within a certain range. Compared with Xception, ShuffleNetV2 and other models used for micro-Doppler signal classification in recent years, the overall performance of ECM Net achieves the best average testing accuracy, model parameter amount and computational complexity. At the same time, this paper also designs a P2P attention module that can achieve point-to-point specific activation from three dimensions with almost no increase in computational burden, improving the efficiency of ECM Net operations and effectively enhancing the accuracy of the model. Finally, the superiority of the proposed method was verified through actual dataset, with an accuracy of 97.4% and a time consumption of approximately 0.18 seconds. • ECM Net achieves multi-scale capture of residual features in complex scenarios. • Compared to traditional networks, ECM Net delivers better accuracy, fewer parameters, and reduced computational demands. • The P2P attention module effectively improves the accuracy of the model without almost no increase in computational burden.