在自动驾驶感知系统中视觉传感器与激光雷达是关键的信息来源,但在目前的3D目标检测任务中大部分纯点云的网络检测能力都优于图像和激光点云融合的网络,现有的研究将其原因总结为图像与雷达信息的视角错位以及异构特征难以匹配,单阶段...在自动驾驶感知系统中视觉传感器与激光雷达是关键的信息来源,但在目前的3D目标检测任务中大部分纯点云的网络检测能力都优于图像和激光点云融合的网络,现有的研究将其原因总结为图像与雷达信息的视角错位以及异构特征难以匹配,单阶段融合算法难以充分融合二者的特征.为此,本文提出一种新的多层多模态融合的3D目标检测方法:首先,前融合阶段通过在2D检测框形成的锥视区内对点云进行局部顺序的色彩信息(Red Green Blue,RGB)涂抹编码;然后将编码后点云输入融合了自注意力机制上下文感知的通道扩充PointPillars检测网络;后融合阶段将2D候选框与3D候选框在非极大抑制之前编码为两组稀疏张量,利用相机激光雷达对象候选融合网络得出最终的3D目标检测结果.在KITTI数据集上进行的实验表明,本融合检测方法相较于纯点云网络的基线上有了显著的性能提升,平均mAP提高了6.24%.展开更多
Fusarium head blight (FHB) is a destructive disease of wheat and other cereals. FHB occurs in Europe, North America and around the world causing significant losses in production and endangers human and animal health. ...Fusarium head blight (FHB) is a destructive disease of wheat and other cereals. FHB occurs in Europe, North America and around the world causing significant losses in production and endangers human and animal health. In this article, we provide the strategic steps for the specific target selection for the phytopathogen system wheat-Fusarium graminearum. The economic impact of FHB leads to the need for innovation. Currently used fungicides have been shown to be effective over the years, but recently cereal infecting Fusaria have developed resistance. Our work presents a new perspective on target selection to allow the development of new fungicides. We developed an innovative approach combining both genomic analysis and molecular modeling to increase the discovery for new chemical compounds with both safety and low environmental impact. Our protein targets selection revealed 13 candidates with high specificity, essentiality and potentially assayable with a favorable accessibility to drug activity. Among them, three proteins: trichodiene synthase, endoglucanase-5 and ERG6 were selected for deeper structural analyses to identify new putative fungicides. Overall, the bioinformatics filtering for novel protein targets applied for agricultural purposes is a response to the demand for chemical crop protection. The availability of the genome, secretome and PHI-base allowed the enrichment of the search that combined experimental data in planta. The homology modeling and molecular dynamics simulations allowed the acquisition of three robust and stable conformers. From this step, approximately ten thousand compounds have been virtually screened against three candidates. Forty-five top-ranked compounds were selected from docking results as presenting better interactions and energy at the binding pockets and no toxicity. These compounds may act as inhibitors and lead to the development of new fungicides.展开更多
针对目前超声3D识别普遍存在的识别率低、鲁棒性差等问题,以物体内部人工标准缺陷为超声靶标,通过对超声靶标脉冲超声回波信号进行处理,提取了相对能量、相对幅值、相对频域带宽、相对峰度系数、相对离散系数、相对包络面积、相对偏度...针对目前超声3D识别普遍存在的识别率低、鲁棒性差等问题,以物体内部人工标准缺陷为超声靶标,通过对超声靶标脉冲超声回波信号进行处理,提取了相对能量、相对幅值、相对频域带宽、相对峰度系数、相对离散系数、相对包络面积、相对偏度系数和相对频谱半高宽等多个特征参数,利用Fisher线性判别分析(Fisher Linear Discriminative Analysis,FLDA)对这些特征参数进行融合,形成融合特征,并采用反向传播(Back Propagation,BP)神经网络对融合特征进行训练与识别,对物体内部矩形槽、横通孔及平底孔三类超声靶标进行识别。试验结果表明:三种靶标的识别率分别高达了93.3%,93.3%,100%;对噪声有抑制能力,对测试工况不敏感,识别稳健性得到了提高,可为超声3D目标识别提供理论和技术参考。展开更多
文摘在自动驾驶感知系统中视觉传感器与激光雷达是关键的信息来源,但在目前的3D目标检测任务中大部分纯点云的网络检测能力都优于图像和激光点云融合的网络,现有的研究将其原因总结为图像与雷达信息的视角错位以及异构特征难以匹配,单阶段融合算法难以充分融合二者的特征.为此,本文提出一种新的多层多模态融合的3D目标检测方法:首先,前融合阶段通过在2D检测框形成的锥视区内对点云进行局部顺序的色彩信息(Red Green Blue,RGB)涂抹编码;然后将编码后点云输入融合了自注意力机制上下文感知的通道扩充PointPillars检测网络;后融合阶段将2D候选框与3D候选框在非极大抑制之前编码为两组稀疏张量,利用相机激光雷达对象候选融合网络得出最终的3D目标检测结果.在KITTI数据集上进行的实验表明,本融合检测方法相较于纯点云网络的基线上有了显著的性能提升,平均mAP提高了6.24%.
文摘Fusarium head blight (FHB) is a destructive disease of wheat and other cereals. FHB occurs in Europe, North America and around the world causing significant losses in production and endangers human and animal health. In this article, we provide the strategic steps for the specific target selection for the phytopathogen system wheat-Fusarium graminearum. The economic impact of FHB leads to the need for innovation. Currently used fungicides have been shown to be effective over the years, but recently cereal infecting Fusaria have developed resistance. Our work presents a new perspective on target selection to allow the development of new fungicides. We developed an innovative approach combining both genomic analysis and molecular modeling to increase the discovery for new chemical compounds with both safety and low environmental impact. Our protein targets selection revealed 13 candidates with high specificity, essentiality and potentially assayable with a favorable accessibility to drug activity. Among them, three proteins: trichodiene synthase, endoglucanase-5 and ERG6 were selected for deeper structural analyses to identify new putative fungicides. Overall, the bioinformatics filtering for novel protein targets applied for agricultural purposes is a response to the demand for chemical crop protection. The availability of the genome, secretome and PHI-base allowed the enrichment of the search that combined experimental data in planta. The homology modeling and molecular dynamics simulations allowed the acquisition of three robust and stable conformers. From this step, approximately ten thousand compounds have been virtually screened against three candidates. Forty-five top-ranked compounds were selected from docking results as presenting better interactions and energy at the binding pockets and no toxicity. These compounds may act as inhibitors and lead to the development of new fungicides.
文摘针对目前超声3D识别普遍存在的识别率低、鲁棒性差等问题,以物体内部人工标准缺陷为超声靶标,通过对超声靶标脉冲超声回波信号进行处理,提取了相对能量、相对幅值、相对频域带宽、相对峰度系数、相对离散系数、相对包络面积、相对偏度系数和相对频谱半高宽等多个特征参数,利用Fisher线性判别分析(Fisher Linear Discriminative Analysis,FLDA)对这些特征参数进行融合,形成融合特征,并采用反向传播(Back Propagation,BP)神经网络对融合特征进行训练与识别,对物体内部矩形槽、横通孔及平底孔三类超声靶标进行识别。试验结果表明:三种靶标的识别率分别高达了93.3%,93.3%,100%;对噪声有抑制能力,对测试工况不敏感,识别稳健性得到了提高,可为超声3D目标识别提供理论和技术参考。
