摘要
基于双目视觉的立体匹配网络被广泛应用于目标定位,其计算速度和精度直接影响机器人的任务执行。针对现有立体匹配网络参数冗余、计算速度缓慢等问题,提出利用混合空洞卷积替代特征提取部分的普通卷积层,减少网络层数,极大地降低了网络参数量,提升了网络的运行速度。为增强网络在弱纹理、边缘区域的特征表达能力,提出一种多尺度感受野模块,以融合不同尺度感受野的特征。将该模块嵌入立体匹配网络,构建了多尺度感受野立体匹配网络,实现了目标位置信息的精确检测。基于所提网络和校正模型,在水池模拟环境下开展了水下机器人的目标定位实验。结果表明,所提方法有效解决了现有立体匹配网络在弱纹理、重叠以及边缘区域的误匹配问题,对特定目标物定位精度可达89%,较其他主流视觉测量方法均有显著提升,在水下目标定位中具有重要应用价值。
Stereo-matching network based on binocular vision has been widely used in target localization,and its calculation speed and accuracy directly affect the task execution of robots.Aiming at the problems of parameter redundancy and slow calculation speed of the existing stereo-matching network,a hybrid dilated convolution is proposed to replace the ordinary convolution layer in the feature extraction part,reduce the number of network layers and parameters and improve the running speed of the network as well.In order to enhance the feature expression ability of the network in weak texture and edge regions,a multi-scale receptive field module is proposed to fuse the features of different scales of receptive fields.This module is embedded into the stereo-matching network,and a multi-scale receptive field stereo-matching network is proposed to realize the accurate detection of target position information.Based on the proposed network and the calibration model,the target localization experiment of the underwater vehicle in the pool simulation environment is carried out.The results show that the proposed method effectively solves the mismatching problem of the existing stereo-matching network in weak texture,overlapping and edge areas,and the localization accuracy of the specific target can reach 89%,which is significantly improved compared with other mainstream vision measurement methods.It has important application value in underwater target location.
作者
涂天佳
秦毅
梁晨
陈然
TU Tianjia;QIN Yi;LIANG Chen;CHEN Ran(College of Mechanical and Vehicle Engineering,Chongqing University,Chongqing 400044,China;State Key Laboratory of Mechanical Transmissions,Chongqing University,Chongqing 400044,China;UAV Department,China Helicopter Research and Development Institute,Jingdezhen 333000,China)
出处
《兵器装备工程学报》
北大核心
2025年第8期11-19,共9页
Journal of Ordnance Equipment Engineering
关键词
双目视觉
深度估计
立体匹配
水下机器人
混合空洞卷积
多尺度
binocular vision
depth estimation
stereo matching
underwater robot
mixed dilated convolution
multiscale
