视觉感知启发的面向出舱活动的物体识别技术研究  

Research on Object Detection Technology Inspired by Visual Perception Theory for EVA

作　　者：张菊莉[1] 马钟 贺占庄[1] 周革强[2] 何双亮[2] 

机构地区：[1]西安微电子技术研究所,西安710065 [2]中国航天员科研训练中心,北京100094

出　　处：《载人航天》2018年第1期41-47,共7页Manned Spaceflight

基　　金：载人航天预先研究项目(060601);中国博士后科学基金资助项目(166553)

摘　　要：为提高航天员出舱活动(EVA)的工效,提出了一种基于视觉感知启发的物体识别方法。首先对视觉观察到的一定区域内的图像进行采集,然后进行二值化赋范梯度的特征提取,并预测物体所在区域的矩形框,选取比该矩形框扩大一定范围的图像作为输入,传递给深度卷积神经网络CNN进行类别识别和精定位。在自建的数据集上进行测试验证,结果表明:该方法达到了88.2%的平均识别准确率,识别速率为0.047 s,可以满足舱外物体识别需求。该方法可为信息化、智能化的出舱活动任务提供参考,对提高出舱活动任务的工效具有重要意义。To improve the efficiency of the extravehicular activity(EVA),an approach for object detection inspired by visual perceptual theory was proposed.First,certain area of the image was collected from visual observation,then the feature maps of binarized normed gradient were extracted,and the bounding box of the object was predicted.Then an area of the image bigger than the bounding box was chosen to feed into the deep convolutional neural networks CNN as the input for class recognition and precise object location.Experiments were conducted to verify the proposed method with the dataset established by ourselves.The results showed that the m AP of this approach was 88.2% and the average compute time was 0.047 s,which could satisfy the demands of EVA.This approach may serve as a reference for the intellectualization and informatization of EVA tasks.It is of great significance for the improvement of EVA efficiency.

关 键 词：出舱活动 视觉感知 物体识别 二值化赋范梯度 深度卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]