融合优选图案的深度学习目标识别及定位技术  

Target recognition and location using deep learning based on selected patterns

作　　者：王立鹏[1] 张智[1] 苏丽[1] 聂文昌 WANG Lipeng;ZHANG Zhi;SU Li;NIE Wenchang(College of Automation,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2020年第4期549-555,共7页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(61803116);中央高校基本科研业务专项资金资助项目(3072019CFJ0405,3072020CF0410);船舶态势智能感知系统研制项目(MC-201920-X01).

摘　　要：为提高复杂环境下视觉系统目标识别和定位精度,本文以深度学习方法为基础,结合辅助图案,提出基于双目视觉系统的卷积神经网络目标识别及定位方法。构建基于Faster Rcnn的网络学习框架,结合具体问题,确定ZF及RPN网络参数;提出图案可分类性概念及其量化评价指标,制定图案优选策略并确定优选图案集;将双目视觉与基于优选图案的深度学习方法相结合,设计复杂环境下三维空间目标识别与定位方法。开展多工况的实际试验。结果表明:本文算法具有较高的识别物体及定位位姿的能力,且具有较好的实用性和鲁棒性。A target recognition and location method that is combined with binocular vision and convolution neural network is proposed in this study. The method is based on deep learning and selected patterns and aims to improve the accuracy of target recognition and location of visual systems under complicated environments. First, a deep learning framework, which uses Faster R-CNN as its core, is established. The parameters of the ZFNet and RPNet are determined on the basis of the specific research target. Second, the concept of pattern classification is proposed, and evaluation criteria are defined to quantify pattern classification. Selected strategies are proposed for the patterns, and a selected pattern assembly is acquired. Third, a three-dimensional space target recognition and location method is designed under complicated environments and combined with binocular vision and deep learning on the basis of the selected patterns. Lastly, the algorithm used in this study is verified through several experiments with different initial conditions. Results show that the proposed method can recognize targets and locate positions and attitudes excellently. Moreover, it has better practicability and robustness compared with other methods.

关 键 词：深度学习 图案分类 辅助图案 优选图案 双目视觉 卷积神经网络 目标识别 定位 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]