A Dual CNN Method for Speed Sensing of Sonar Robots

To improve the perception accuracy of sonar robots' speed, this paper introduces a dual-CNN approach based on spectrum and time-frequency features. Rough features with high-speed and fine features with low-speed are separately extracted through spectral and time-frequency map. Then, these features are put into two CNN networks for training, and weighted voting method is finally applied to obtain the classification results. Experimental results demonstrate that the dual-CNN’s perceptual accuracy is improved by 1% to 6% compared to the single CNN model. Furthermore, the inclusion of rough and fine features enhances feature diversity, thereby enhancing the speed perception accuracy and stability of sonar robots.