BADE: BEV-Adaptive Distance Estimation Neural Network based on Monocular Cameras

Abstract

Distance estimation is a pivotal component of environment perception of autonomous driving. The current mainstream distance estimation methods are mainly based on binocular cameras and laser radars, which are, however, either difficult to calibrate or expensive. It is promising to perform distance estimation based on monocular cameras, which are less expensive and easier to calibrate. But depth information is lost in a single image of monocular cameras so that it is more challenging to ensure the distance estimation precision compared with other methods. To address this challenge, this paper proposes a neural network for monocular distance estimation, which estimates the distance of objects within a single image in an end-to-end way. More specifically, we implement an object detection neural network to detect objects, fuse that object detection network and the distance estimation methods based on bird's-eye view to form our BEV-Adaptive Distance Estimation Neural Network(BADE) based on bird's-eye view, which generates multi-objective distance estimation. Meanwhile, we enhance our BADE by inserting an auxiliary network branch into the object detection network, which takes the distance estimation results as feedback to better train the object detection neural network and can improve the distance estimation precision. Experiments demonstrate the performance of the proposed BADE.