基于单目深度估计和校准参数的距离测算方法  

Distance measurement method based on monocular depth estimation and calibration parameters

作　　者：余萍[1] 胡旭欣 Yu Ping;Hu Xuxin(North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学,保定071003

出　　处：《电子测量技术》2022年第20期88-94,共7页Electronic Measurement Technology

摘　　要：为了提升有监督学习的单目深度估计网络对于实际场景测距任务的准确性和适用性,提出了一种基于单目深度估计和校准参数的距离测算方法。首先通过引入多元注意力模块和优化设计网络结构,构建了一种融合全局上下文和空间注意力机制的网络(GSNet),然后制定校准参数以建立场景的预测距离与实际距离的比例关系,从而获得校准后的距离值。实验证明,融合网络GSNet和校准参数可以有效减小单目深度估计方法在实际测算距离的误差。相比于使用单目深度估计直接预测距离信息,本文方法测算距离的平均绝对误差小于0.15 m,平均相对误差小于10%,具有很好的可行性和准确性。To improve the accuracy and applicability of the monocular depth estimation network with supervised learning for actual scene ranging tasks, a distance calculation method based on monocular depth estimation and calibration mechanism is proposed. Firstly, by introducing multivariate attention blocks and optimizing the design network structure, a network integrating global context and spatial attention mechanism(GSNet) is constructed. Then calibration parameters are formulated to establish the proportional relationship between the predicted distance of the scene and the actual distance of the scene, to obtain the calibrated distance value. Experimental results show that the fusion network GSNet and calibration mechanism can effectively reduce the error of the monocular depth estimation method in the actual measured distance. Compared with other methods, the average absolute error is less than 0.15 m, and the average relative error of the measured distance in this method is less than 10%, which has good feasibility and accuracy.

关 键 词：单目深度估计 卷积神经网络 全局上下文模块 空间注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]