Wideband DOA Estimation Based on ConvNeXt
Xinya Li, Chunjie Zhang · 2024
This paper proposes a ConvNeXt method for direction of arrival (DOA) estimation aimed at enhancing the accuracy of wideband signals. The method utilizes the spatial covariance matrix components of each frequency subband as inputs to the network, thereby leveraging the data characteristics across different frequency subbands of the wideband signal to facilitate the regression of angular values, it uses DOA estimation as a regression task to directly obtain DOA values. Additionally, the CosineAnnealing algorithm is employed during the training process to adaptively adjust the network’s learning rate to improve the effect. The experimental results prove the effectiveness of the proposed method. The ConvNeXt network exhibits superior estimation error performance compared to traditional methods, particularly at low signal-to-noise ratio (SNR) and limited snapshots. Furthermore, this method maintains low algorithmic complexity while ensuring high computational efficiency.