DeepRange: deep-learning-based object detection and ranging in autonomous driving

摘要

Autonomous driving is an emerging area of intelligent transport systems. It necessitates automatic detection, classification, and ranging of on-road obstacles. Current autonomous driving systems rely largely on LiDAR and radar units to gather information of distance to obstacles. LiDAR units are, in general, expensive. Alternatives such as stereo image processing for obtaining distance estimates are computationally intensive. Here, the authors explore the power of a deep-learning-based approach for range finding. The proposed approach is based on perception and requires only a low-cost image sensor. Estimating the range of objects from a monocular image captured by a passive cost-effective image sensor is, however, a challenging task. The authors propose an enhancement to classical convolutional neural networks based on addition of a range estimation layer for obtaining the distance to detected objects. The proposed strategy accomplishes object detection, classification and ranging simultaneously. The approach has been studied on the KITTI Vision Benchmark Suite, and experimental results indicate a detection speed of 61 fps, with mAP of 96.92% in recognition performance on an NVIDIA RTX 2080Ti GPU platform. Further, the proposed approach leads to an average error of only 0.915 m in range estimation which is quite acceptable in highway scenarios.