In dynamic scenes,the pose estimation and map consistency of visual simultaneous localisation and mapping(visual SLAM)are affected by intermittent changes in object motion states.An adaptive motion-state estimation an...In dynamic scenes,the pose estimation and map consistency of visual simultaneous localisation and mapping(visual SLAM)are affected by intermittent changes in object motion states.An adaptive motion-state estimation and feature-reuse mechanism is proposed which restores features once objects become stationary.Camera ego-motion is com-pensated via projection-based point-to-point red-green-blue-depth(RGB-D)Iterative Closest Point;the alignment residual yields a short-term jitter score.An Extended Kalman Filter fuses the centre-pixel trajectory and depth of the object,using depth innovation as strong evidence to suppress false triggers.Applied adaptive decision thresholds involve resolution,ego-motion intensity,jitter,and reference depth,and are combined with dual/single triggering and hysteresis to achieve robust switching.When an object is considered static,its feature points are reused.On the Bonn RGB-D Dynamic Dataset(BONN)and TUM RGB-D SLAM Dataset and Benchmark(TUM),the proposed method matches or exceeds baselines:In intermittent-motion-dominated BONN sequences Placing_non_box,it re-duces the root-mean-square of the absolute trajectory error(ATE-RMSE)by 27%relative to the baseline,remains comparable to Ellipsoid-SLAM on TUM,and consistently outperforms ORB-SLAM3 in dynamic scenes.The hysteresis counter reading on Placing_non_box2 shows that the proposed method can reduce the motion-state misclassification rate by nearly 40%.From the ablation experiment results,we confirm that adaptive thresholds yield the most significant optimisation effect.The approach improves robustness and map completeness in dynamic environments without degrading performance in low-dynamic settings.展开更多
针对卷积神经网络(CNN)在感受野有限、缺乏对全局信息的有效感知,以及在处理短时稳态运动视觉诱发电位(SSMVEP)信号时分类效果欠佳的问题,提出了一种紧凑EEGNet-Transformer(即EEGNetformer)网络。EEGNetformer网络融合了为脑电(EEG)信...针对卷积神经网络(CNN)在感受野有限、缺乏对全局信息的有效感知,以及在处理短时稳态运动视觉诱发电位(SSMVEP)信号时分类效果欠佳的问题,提出了一种紧凑EEGNet-Transformer(即EEGNetformer)网络。EEGNetformer网络融合了为脑电(EEG)信号识别任务而设计的通用的卷积神经网络EEGNet网络和Transformer网络的优势,有效地捕捉与处理脑电信号中的局部和全局信息,增强网络对SSMVEP特征的学习,进而实现良好的解码性能。EEGNet网络用于提取SSMVEP的局部时间和空间特征,而Transformer网络用于捕捉脑电时间序列的全局信息。在基于SSMVEP-BCI范式采集的数据基础上,开展了实验以评估EEGNetformer网络的性能。实验结果显示,当在2 s SSMVEP数据条件下,EEGNetformer网络在基于被试者内情况的平均准确率为88.9%±6.6%,在基于跨被试者情况的平均准确率为69.1%±4.3%。与传统的CNN算法相比,EEGNetformer网络的分类性能提升了4.2%~17.4%。研究内容说明,EEGNetformer网络在有效提高SSMVEP-BCI识别准确率方面具有显著优势,为进一步提升SSMVEP-BCI解码性能提供了新的研究思路。展开更多
文摘In dynamic scenes,the pose estimation and map consistency of visual simultaneous localisation and mapping(visual SLAM)are affected by intermittent changes in object motion states.An adaptive motion-state estimation and feature-reuse mechanism is proposed which restores features once objects become stationary.Camera ego-motion is com-pensated via projection-based point-to-point red-green-blue-depth(RGB-D)Iterative Closest Point;the alignment residual yields a short-term jitter score.An Extended Kalman Filter fuses the centre-pixel trajectory and depth of the object,using depth innovation as strong evidence to suppress false triggers.Applied adaptive decision thresholds involve resolution,ego-motion intensity,jitter,and reference depth,and are combined with dual/single triggering and hysteresis to achieve robust switching.When an object is considered static,its feature points are reused.On the Bonn RGB-D Dynamic Dataset(BONN)and TUM RGB-D SLAM Dataset and Benchmark(TUM),the proposed method matches or exceeds baselines:In intermittent-motion-dominated BONN sequences Placing_non_box,it re-duces the root-mean-square of the absolute trajectory error(ATE-RMSE)by 27%relative to the baseline,remains comparable to Ellipsoid-SLAM on TUM,and consistently outperforms ORB-SLAM3 in dynamic scenes.The hysteresis counter reading on Placing_non_box2 shows that the proposed method can reduce the motion-state misclassification rate by nearly 40%.From the ablation experiment results,we confirm that adaptive thresholds yield the most significant optimisation effect.The approach improves robustness and map completeness in dynamic environments without degrading performance in low-dynamic settings.
文摘针对卷积神经网络(CNN)在感受野有限、缺乏对全局信息的有效感知,以及在处理短时稳态运动视觉诱发电位(SSMVEP)信号时分类效果欠佳的问题,提出了一种紧凑EEGNet-Transformer(即EEGNetformer)网络。EEGNetformer网络融合了为脑电(EEG)信号识别任务而设计的通用的卷积神经网络EEGNet网络和Transformer网络的优势,有效地捕捉与处理脑电信号中的局部和全局信息,增强网络对SSMVEP特征的学习,进而实现良好的解码性能。EEGNet网络用于提取SSMVEP的局部时间和空间特征,而Transformer网络用于捕捉脑电时间序列的全局信息。在基于SSMVEP-BCI范式采集的数据基础上,开展了实验以评估EEGNetformer网络的性能。实验结果显示,当在2 s SSMVEP数据条件下,EEGNetformer网络在基于被试者内情况的平均准确率为88.9%±6.6%,在基于跨被试者情况的平均准确率为69.1%±4.3%。与传统的CNN算法相比,EEGNetformer网络的分类性能提升了4.2%~17.4%。研究内容说明,EEGNetformer网络在有效提高SSMVEP-BCI识别准确率方面具有显著优势,为进一步提升SSMVEP-BCI解码性能提供了新的研究思路。
