Layer Selection for Subtraction and Concatenation: A Method for Visual Velocity Estimation of a Mobile Robot

Abstract

Kinematic information such as position, velocity, and acceleration is critical to determine the three-dimensional state of the robot in space. In this study, it is aimed at estimating, visually, the linear and angular velocity of a mobile robot. Additionally, another aim of this study is to determine the suitability of the concatenation or subtraction layer in the Convolutional Neural Network (CNN) that will make this estimate. For these purposes, first, a simulation environment was created. 9000 pairs of images and the necessary velocity information were collected from this simulation environment for training. Similarly, 1000 pairs of images and velocity information were gathered for validation. Four different CNN models were designed, and these models were trained and tested using these datasets. As a result of the test, the lowest average error for linear velocity estimation was calculated as 0.93e-3m/s, and the angular velocity estimation was measured as 4.37e-3rad/s. It was observed that the results were sufficient for linear and angular velocity prediction, according to the statistical analysis of errors. In addition, it was observed that the subtraction layer can be used instead of the concatenation layer in the CNN architectures for hardware-limited systems. As a result, visual velocity estimation of mobile robots has been achieved with this study, and the framework of CNN models has been drawn for this problem.