Energy-efficient time-of-flight estimation in the presence of outliers: A machine learning approach

The time-of-flight (ToF) estimation problem is common in sonar, ultrasound, radar, and other remote sensing applications. The conventional ToF maximum-likelihood estimator (MLE) exhibits a rapid deterioration in the accuracy when the signal-to-noise ratio (SNR) falls below a certain threshold. This threshold effect emerges mostly due to appearance of outliers associated with the side lobes in the autocorrelation function of a narrowband source signal. In our previous work, we have introduced a bank of unmatched filters and biased ToF estimators derived using these filters. These biased estimators form a feature vector for training a classifier which, subsequently, is used for reducing the bias and the variance parts induced by outliers in the mean-square error (MSE) of the MLE. In this paper, we extend the above method by introducing an adaptive scheme for controlling the number of measurements (pulses) required to achieve a desired accuracy. We show that using the information provided by a classifier, it is possible to achieve the estimation error of the MLE but by using significantly less number of pulses and thus energy on average.