Automotive Radar based Road Boundary Estimation using a light-weight Regression CNN

In this paper, we consider the problem of estimating the road boundary along the path of an autonomous vehicle using automotive radar sensor measurements. To do this, we propose a light-weight convolutional neural network (CNN) that uses the radar image depicting the bird eye view of the front environment of the vehicle as input. The road boundaries to the left and right are represented as a collection of key points that are regularly placed in longitudinal direction. The output layer of the regression CNN is designed to predict the lateral coordinates of these key points. The proposed light-weight CNN architecture, designed specifically for the radar images which are sparse in nature, is compared with the well-known architectures like Xception, VGG-16 etc. It is found that the root mean square error (RMSE) and $R^{2}$ score performance of the proposed model is comparable to Xception model, but with hardware complexity reduced by 100 times. Also, to reduce the manual labeling effort, we have devised a novel and reliable auto-labeling approach to derive the road boundary ground truths.