A study on estimation of car orientation using convolution neural networks

Abstract

While advanced driver assistance system becomes popular in these days, further improvement of detection and recognition accuracy of cars using in-vehicle cameras is required. If it is possible to estimate the orientation of the car in the image, the direction of the car movement will be able to be predicted and be helpful for developing safer advanced driving support systems. In this study, we propose an estimation method for car orientation in the image using convolutional neural networks. The proposed method employed image pre-filtering processes for sharpening the images and both of batch normalization and dropout for preventing gradient loss and over-fitting in the learning procedure.