Estimation of vehicle dynamics by fusion image and radar based on subtraction convolutional neural network

Abstract

Abstract The main goal of this paper is to estimate the vehicle dynamic by image and radar fusion through a deep learning model. In the designed deep learning model, multiple convolutional neural networks are used. In addition, tensor subtraction is added in the model to express the time-series features. In addition to the fusion of images and radar, the accuracy of using a single sensor to estimate vehicle dynamics is also tested. The result shows that the performance of using the fusion of radar and image is better than using a single sensor. Finally, by using both signals in the deep learning structure, the mean square errors of estimation angular x, angular y, angular z, and linear velocity obtained from our model have values of 2.86e-6, 4.72e-6, 6.19e-5, and 2.41e-2, respectively.