Convolutional Gated Recurrent Units for Obstacle Segmentation in Bird-Eye-View

Abstract

Obstacle detection is a fundamental problem in autonomous driving. The most common solutions share the idea of modeling the free-space and marking as obstacles all the points that lie outside this model according to a threshold. Manually setting this threshold and adapting the model to the various scenarios is not ideal, whereas a machine learning approach is more suitable for this kind of task. In this work we present an application of Convolutional Neural Networks (CNNs) for the detection of obstacles in front of a vehicle. Our goal is to train a CNN to understand which patterns in this area are connected to the presence of obstacles. Our method does not require any manual annotation, since the training relies on a classification that comes from a LiDAR. During inference, our network requires as input a 3D point cloud generated from stereoscopic images. Moreover, we make use of recurrent units in our network, since they are able to exploit temporal information to provide more accurate results in case of occlusion. We compare different input configurations and show that our final selection is able to correctly predict the position of obstacles and to generalize well in unseen environments.