目前国内尚无公开的多视角手语数据集,手语识别研究主要围绕单一视角数据展开,由于手势遮挡问题,模型识别效果不佳.针对这些问题,创建了一个多视角孤立手语数据集(Multi-View Chinese Isolated Sign Language Dataset,MV-CISL);基于该...目前国内尚无公开的多视角手语数据集,手语识别研究主要围绕单一视角数据展开,由于手势遮挡问题,模型识别效果不佳.针对这些问题,创建了一个多视角孤立手语数据集(Multi-View Chinese Isolated Sign Language Dataset,MV-CISL);基于该数据集,提出了一种多视角特征融合的孤立手语识别方法,该方法使用基于改进的3D-ResNet18的端到端多流网络提取不同视角的特征信息,并通过决策级融合来整合这些特征信息;为提高网络识别性能,使用CSL-500单视角数据集对所提出网络进行迁移学习,并将其应用于MVCISL数据集.实验结果表明,所提出方法在性能上优于单视角和双视角方法;在多流网络骨干模型ResNet+LSTM、ResNet+BiLSTM、3D-MobileNet和3D-ShuffleNet上进一步验证了该方法的有效性;与基于正面视角RGB和深度信息融合的方法相比,数据采集成本更低,性能更优良.展开更多
针对动态环境中实时定位与建图(Simultaneous Localization and Mapping,SLAM)算法位姿估计存在的定位漂移、实时性差等问题,提出一个名为YSG-SLAM的实时语义RGB-D SLAM系统。为了提高系统实时性,新增两个并行线程:一个用于获取二维语...针对动态环境中实时定位与建图(Simultaneous Localization and Mapping,SLAM)算法位姿估计存在的定位漂移、实时性差等问题,提出一个名为YSG-SLAM的实时语义RGB-D SLAM系统。为了提高系统实时性,新增两个并行线程:一个用于获取二维语义信息的语义分割线程,一个语义建图线程。为优化系统在处理动态物体时的准确性和鲁棒性,YSG-SLAM引入快速动态特征剔除算法,并耦合漏检补偿模块来应对基于实时实例分割(You Only Look At Coefficients,YOLACT)算法可能出现的漏检情况,有效地提升了特征点剔除的精确度和系统的整体稳定性。为减少由特征点聚集引起的定位误差从而优化特征点的空间分布,设计自适应角点提取阈值计算方法,使特征分布更加均匀。语义建图线程充分利用二维语义信息与三维点云数据,可选择性构建语义地图和八叉树地图,提高了系统的环境感知能力及机器人在复杂环境下的相关任务执行能力。YSG-SLAM在德国慕尼黑工业大学数据集、Bonn数据集上进行了评估,相较于原ORB-SLAM2,各项定位误差下降达93%。实验结果表明,YSG-SLAM有效提升了系统实时性,定位精度高,且可构建两种地图,具有一定的实用价值。展开更多
针对同步定位与建图(Simultaneous Localization and Mapping,SLAM)算法在动态场景中存在定位精度低且无法生成稠密地图的问题,提出一种基于动态区域剔除与稠密建图的视觉SLAM算法。在原ORB-SLAM3算法的基础上新建动态特征点检测线程,使...针对同步定位与建图(Simultaneous Localization and Mapping,SLAM)算法在动态场景中存在定位精度低且无法生成稠密地图的问题,提出一种基于动态区域剔除与稠密建图的视觉SLAM算法。在原ORB-SLAM3算法的基础上新建动态特征点检测线程,使用YOLOX网络获取动态场景语义信息及物体检测框,同时结合语义和几何约束检测特征点运动状态,提出动态特征点剔除算法,旨在精准实现动态特征点的剔除。随后设计稠密建图线程,基于关键帧及相应位姿构建稠密点云地图,利用地图中剩余的静态特征点,去除动态物体造成的重影,实现稠密地图的构建。在公开TUM数据集和真实动态环境进行验证,在TUM数据集的动态环境下,新算法有效消除了动态物体对位姿估计的影响,提升了SLAM算法在动态场景中的定位与建图的鲁棒性。展开更多
High-temperature piezoelectric sen-sors are very important in severe environments such as fire safety,aerospace and oil drills,however,most current sensors are not heat res-istant(<300℃)and are fragile,which limit...High-temperature piezoelectric sen-sors are very important in severe environments such as fire safety,aerospace and oil drills,however,most current sensors are not heat res-istant(<300℃)and are fragile,which limits their use,especially in high-temperature environ-ments.A high-temperature resistant flexible piezoelectric film based on graphene oxide(GO)/polyacrylonitrile(PAN)composites was prepared by electrospinning and thermal treat-ment.It was packed into a micro-device,which could work continuously at 500℃.The intro-duction of GO significantly increased the mechanical properties of the PAN nanofibers because the oxygen-containing func-tional groups(electronegative groups)on the surface of the GO initiated a nucleophilic attack on the PAN molecule during heat treatment,enabling the GO to initiate the cyclization of the PAN at lower heat-treatment temperatures.In addition,the abund-ant oxygen-containing functional groups on GO acted as pro-oxidants to hasten the oxidation of PAN during heat treatment.The effects of GO content and heat treatment temperature on the properties of the nanofiber films were investigated.A GO/PAN nanofiber piezoelectric sensor heat-treated at 300℃had a 9.10 V and 2.25μA peak output,which are respectively 101.3%and 78.6%higher than those of the untreated films.Cyclic testing over 5000 cycles at 350℃confirmed the stable out-put performance of the GO/PAN nanofiber piezoelectric sensor.Furthermore,a sensor heat-treated at 400℃had a sensitivity of 1.7 V/N,which is 83.5%higher than that of an untreated one.The results show that the prepared GO/PAN nanofiber piezo-electric sensor combines high temperature resistance,high flexibility,stability and high sensitivity,and may have broad applic-ations in high temperature environments such as the aerospace and petroleum industries.展开更多
文摘目前国内尚无公开的多视角手语数据集,手语识别研究主要围绕单一视角数据展开,由于手势遮挡问题,模型识别效果不佳.针对这些问题,创建了一个多视角孤立手语数据集(Multi-View Chinese Isolated Sign Language Dataset,MV-CISL);基于该数据集,提出了一种多视角特征融合的孤立手语识别方法,该方法使用基于改进的3D-ResNet18的端到端多流网络提取不同视角的特征信息,并通过决策级融合来整合这些特征信息;为提高网络识别性能,使用CSL-500单视角数据集对所提出网络进行迁移学习,并将其应用于MVCISL数据集.实验结果表明,所提出方法在性能上优于单视角和双视角方法;在多流网络骨干模型ResNet+LSTM、ResNet+BiLSTM、3D-MobileNet和3D-ShuffleNet上进一步验证了该方法的有效性;与基于正面视角RGB和深度信息融合的方法相比,数据采集成本更低,性能更优良.
文摘针对动态环境中实时定位与建图(Simultaneous Localization and Mapping,SLAM)算法位姿估计存在的定位漂移、实时性差等问题,提出一个名为YSG-SLAM的实时语义RGB-D SLAM系统。为了提高系统实时性,新增两个并行线程:一个用于获取二维语义信息的语义分割线程,一个语义建图线程。为优化系统在处理动态物体时的准确性和鲁棒性,YSG-SLAM引入快速动态特征剔除算法,并耦合漏检补偿模块来应对基于实时实例分割(You Only Look At Coefficients,YOLACT)算法可能出现的漏检情况,有效地提升了特征点剔除的精确度和系统的整体稳定性。为减少由特征点聚集引起的定位误差从而优化特征点的空间分布,设计自适应角点提取阈值计算方法,使特征分布更加均匀。语义建图线程充分利用二维语义信息与三维点云数据,可选择性构建语义地图和八叉树地图,提高了系统的环境感知能力及机器人在复杂环境下的相关任务执行能力。YSG-SLAM在德国慕尼黑工业大学数据集、Bonn数据集上进行了评估,相较于原ORB-SLAM2,各项定位误差下降达93%。实验结果表明,YSG-SLAM有效提升了系统实时性,定位精度高,且可构建两种地图,具有一定的实用价值。
文摘针对同步定位与建图(Simultaneous Localization and Mapping,SLAM)算法在动态场景中存在定位精度低且无法生成稠密地图的问题,提出一种基于动态区域剔除与稠密建图的视觉SLAM算法。在原ORB-SLAM3算法的基础上新建动态特征点检测线程,使用YOLOX网络获取动态场景语义信息及物体检测框,同时结合语义和几何约束检测特征点运动状态,提出动态特征点剔除算法,旨在精准实现动态特征点的剔除。随后设计稠密建图线程,基于关键帧及相应位姿构建稠密点云地图,利用地图中剩余的静态特征点,去除动态物体造成的重影,实现稠密地图的构建。在公开TUM数据集和真实动态环境进行验证,在TUM数据集的动态环境下,新算法有效消除了动态物体对位姿估计的影响,提升了SLAM算法在动态场景中的定位与建图的鲁棒性。
文摘High-temperature piezoelectric sen-sors are very important in severe environments such as fire safety,aerospace and oil drills,however,most current sensors are not heat res-istant(<300℃)and are fragile,which limits their use,especially in high-temperature environ-ments.A high-temperature resistant flexible piezoelectric film based on graphene oxide(GO)/polyacrylonitrile(PAN)composites was prepared by electrospinning and thermal treat-ment.It was packed into a micro-device,which could work continuously at 500℃.The intro-duction of GO significantly increased the mechanical properties of the PAN nanofibers because the oxygen-containing func-tional groups(electronegative groups)on the surface of the GO initiated a nucleophilic attack on the PAN molecule during heat treatment,enabling the GO to initiate the cyclization of the PAN at lower heat-treatment temperatures.In addition,the abund-ant oxygen-containing functional groups on GO acted as pro-oxidants to hasten the oxidation of PAN during heat treatment.The effects of GO content and heat treatment temperature on the properties of the nanofiber films were investigated.A GO/PAN nanofiber piezoelectric sensor heat-treated at 300℃had a 9.10 V and 2.25μA peak output,which are respectively 101.3%and 78.6%higher than those of the untreated films.Cyclic testing over 5000 cycles at 350℃confirmed the stable out-put performance of the GO/PAN nanofiber piezoelectric sensor.Furthermore,a sensor heat-treated at 400℃had a sensitivity of 1.7 V/N,which is 83.5%higher than that of an untreated one.The results show that the prepared GO/PAN nanofiber piezo-electric sensor combines high temperature resistance,high flexibility,stability and high sensitivity,and may have broad applic-ations in high temperature environments such as the aerospace and petroleum industries.
