Accurate Off-Grid TOA Estimation Based on Deep Convolution Networks

An accurate off-grid time of arrival (TOA) estimation method is proposed in this paper, utilizing sparse coding-based deep convolutional networks and the matched filtering (MF) algorithm. A rough-accurate estimation structure is constructed in the process of time delay estimation. A correlator processor is contributed to calculate coarse delay values. And a TOA difference regressor driven by sparse coding-based deep convolutional network may then be trained to determine the difference between coarse values and true delay times. The high-precision TOA estimation can be obtained by adding the results of the processor and regressor. Compared with traditional TOA estimation methods based on MF algorithm, the proposed data-driven deep convolutional network can break through the constraint of data sampling rate, accurately estimate the TOA which does not coincide with an integer multiple of the sampling points, and have more accurate estimation performance at lower signal-to-noise levels. In the scenario of TOA estimation of the direct sound wave from the far field, the estimation accuracy of the algorithm proposed in this paper is significantly improved compared to the MF algorithm. When SNR<10dB, the RMSE value of the algorithm decreases by more than 66%.