Research on Depth Perception Algorithm Based on In-
Vehicle Cameras
ABSTRACT
Monocular depth estimation is an important research topic in the field
of computer vision, which can be applied to fields such as autonomous
driving, 3D reconstruction, and virtual reality. In recent years, with the
continuous development of autonomous driving technology, monocular
depth estimation methods suitable for traffic scenes have more application
prospects. This method can estimate the depth information of the scene
from a single image without depth sensors, which can improve the
environmental perception ability of autonomous driving vehicles.
Meanwhile, in depth perception algorithms, self supervised methods are
playing an increasingly important role in monocular depth estimation, with
greater potential and lower annotation costs compared to supervised
methods. Therefore, this article focuses on the problem of self supervised
monocular camera depth estimation in traffic scenes, and the main research
content is as follows:
A dynamic convolution based self supervised monocular depth
estimation algorithm (OE-Depth) is designed to address the issue of
inaccurate and low accuracy prediction results at object edges in existing
self supervised monocular depth estimation algorithms. This algorithm
uses a multi-dimensional dynamic convolutional feature extraction
network to obtain richer feature representations to improve the algorithm's
predictive ability. We also reduce the network's loss at the edges of objects
and optimize the algorithm by adding a triplet loss term. The effectiveness
of the proposed algorithm (OE-Depth) has been demonstrated through
training, validation, and testing on the KITTI dataset. The error class
evaluation index of the algorithm has been optimized by nearly 10%, and
the strictest index in depth prediction δ<1.25, has achieved an accuracy of
