UpBEV: Fast and Accurate LiDAR-Based Drivable Region Detection Utilizing Uniform Polar BEV

Abstract

Drivable region detection is a crucial upstream task for autonomous navigation, so speed and accuracy are the most critical indicators for safe driving. In this article, we proposed a novel representation paradigm for LiDAR data, whereby the drivable region can be efficiently detected and transformed into a dense region in the bird's eye view. Our method differs from the conventional spatial feature extraction and deep learning-based computation-intensive methods. Based on the proposed representation paradigm, our method takes full advantage of image-based features and processing to capture the boundaries between drivable and non-drivable regions within 10 ms solely on a CPU clocked at 4.0 GHz, thus suitable for most mobile platforms with various computational resources. Our contributions are fourfold. Firstly, we propose UpBEV, a representation addressing the sparsity of the point cloud from LiDAR. With this representation, the boundaries are projected into a 2D image and become distinguishable. Second, we develop a complete framework for road detection based on UpBEV, directly generating a dense top-view drivable region that is essential for navigation. Third, with comprehensive experiments on KITTI-Road dataset and SemanticKITTI dataset, the accuracy, speed, and robustness of our method are demonstrated well. Particularly, our method outperforms all the state-of-the-art non-learning methods on the KITTI-Road Benchmark in both maximum F1-measure and runtime, regardless of data type.