Depth Completion in Autonomous Driving: Adaptive Spatial Feature Fusion and Semi-Quantitative Visualization

Abstract

The safety of autonomous driving is closely linked to accurate depth perception. With the continuous development of autonomous driving, depth completion has become one of the crucial methods in this field. However, current depth completion methods have major shortcomings in small objects. To solve this problem, this paper proposes an end-to-end architecture with adaptive spatial feature fusion by encoder–decoder (ASFF-ED) module for depth completion. The architecture is built on the basis of the network architecture proposed in this paper, and is able to extract depth information adaptively with different weights on the specified feature map, which effectively solves the problem of insufficient depth accuracy of small objects. At the same time, this paper also proposes a depth map visualization method with a semi-quantitative visualization, which makes the depth information more intuitive to display. Compared with the currently available depth map visualization methods, this method has stronger quantitative analysis and horizontal comparison ability. Through experiments of ablation study and comparison, the results show that the method proposed in this paper exhibits a lower root-mean-squared error (RMSE) and better small object detection performance on the KITTI dataset.