Bat-inspired velocity estimation neural network in adverse weather conditions

Abstract

When building an autonomous driving system, measuring velocity is essential to predict or track the position of objects. However, cameras do not work properly in adverse weather, making it difficult to measure velocity. In this paper, we deal with a neural network that imitates bats and measures the position and velocity of moving objects even in adverse weather such as sudden fog or rain. Next, a method to stabilize the object location accuracy of the current system using the measured velocity was presented. In the proposed method, a composite Hyperbolic Frequency Modulated signal (cHFM) was presented to extract the Doppler effect from ultrasound signals. Additionally, the proposed network considered velocity profiles during the training phase to measure relative velocity. Lastly, in the proposed object position accuracy stabilization method, actually measured velocity information was added to the Kalman filter. The performance of the proposed network and stabilization method was proven through F1 score, RMSE, and MAE values, which showed that location accuracy can be improved by simultaneously measuring the speed of objects even in adverse weather.