Methods for extracting features from time series data using deep learning have been widely studied,but they still suffer from problems of severe loss of feature information across different network layers and paramete...Methods for extracting features from time series data using deep learning have been widely studied,but they still suffer from problems of severe loss of feature information across different network layers and parameter redun-dancy.Therefore,a new time-series data feature extraction model(CNN-CBAM)that integrates convolutional neural networks(CNN)and convolutional attention mechanisms(CBAM)is proposed.First,the parameters of the CNN and BiGRU prediction models are optimized through uniform design methods.Next,the CNN is used to extract features from the time series data,outputting multiple feature maps.These feature maps are then subjected to feature re-extraction by the CBAM attention mechanism at both the spatial and channel levels.Finally,the feature maps are input into the BiGRU model for prediction.Experimental results show that after CNN-CBAM processing,the stability and accuracy of the BiGRU pre-diction model improved by 77.6%and 76.3%,respectively,outperforming other feature extraction methods.Meanwhile,the training time of the model has only increased by 7.1%,demonstrating excellent time efficiency.展开更多
在视觉同时定位与地图构建问题中,ORB(Oriented FAST and Rotated BRIEF)特征由于其高效、稳定的优点而受到广泛关注。针对ORB特征提取过程中存在的像点量测精度较低、特征聚集现象明显等问题,提出了一种适用于高精度SLAM的均衡化亚像素...在视觉同时定位与地图构建问题中,ORB(Oriented FAST and Rotated BRIEF)特征由于其高效、稳定的优点而受到广泛关注。针对ORB特征提取过程中存在的像点量测精度较低、特征聚集现象明显等问题,提出了一种适用于高精度SLAM的均衡化亚像素ORB特征提取方法。分析了精确特征定位的原理,对误差方程进行合理的简化并采用一种基于模板窗口距离的权函数计算方法,大幅降低了计算负担;设计了一种基于四叉树结构的特征均衡化方案,对包含特征的像平面空间进行有限次数的迭代分割,然后选取具有最优响应的特征。试验表明,本文方法进行特征提取的额外计算负担小于2.5 ms,在运行TUM和KITTI数据集时,ORB特征的量测精度分别为0.84和0.62 Pixel,达到亚像素水平,可以降低误差初值,提高光束法平差效率,并能够在满足特征总体分布规律的情况下,显著改善特征聚集的现象,有利于后续问题的稳健、准确求解。展开更多
文摘Methods for extracting features from time series data using deep learning have been widely studied,but they still suffer from problems of severe loss of feature information across different network layers and parameter redun-dancy.Therefore,a new time-series data feature extraction model(CNN-CBAM)that integrates convolutional neural networks(CNN)and convolutional attention mechanisms(CBAM)is proposed.First,the parameters of the CNN and BiGRU prediction models are optimized through uniform design methods.Next,the CNN is used to extract features from the time series data,outputting multiple feature maps.These feature maps are then subjected to feature re-extraction by the CBAM attention mechanism at both the spatial and channel levels.Finally,the feature maps are input into the BiGRU model for prediction.Experimental results show that after CNN-CBAM processing,the stability and accuracy of the BiGRU pre-diction model improved by 77.6%and 76.3%,respectively,outperforming other feature extraction methods.Meanwhile,the training time of the model has only increased by 7.1%,demonstrating excellent time efficiency.
文摘在视觉同时定位与地图构建问题中,ORB(Oriented FAST and Rotated BRIEF)特征由于其高效、稳定的优点而受到广泛关注。针对ORB特征提取过程中存在的像点量测精度较低、特征聚集现象明显等问题,提出了一种适用于高精度SLAM的均衡化亚像素ORB特征提取方法。分析了精确特征定位的原理,对误差方程进行合理的简化并采用一种基于模板窗口距离的权函数计算方法,大幅降低了计算负担;设计了一种基于四叉树结构的特征均衡化方案,对包含特征的像平面空间进行有限次数的迭代分割,然后选取具有最优响应的特征。试验表明,本文方法进行特征提取的额外计算负担小于2.5 ms,在运行TUM和KITTI数据集时,ORB特征的量测精度分别为0.84和0.62 Pixel,达到亚像素水平,可以降低误差初值,提高光束法平差效率,并能够在满足特征总体分布规律的情况下,显著改善特征聚集的现象,有利于后续问题的稳健、准确求解。
