海上地震拖缆自动定深器建模及仿真研究

本文首先对自行研制的海上地震拖缆自动定深器建立了较为详细的数学模型,给出了其传递函数模型,然后在Matlab环境下,利用仿真工具Simulink和控制系统工具箱（ControlSystemToolbox）中的SISO工具设计了经典控制器,并且对该系统进行了仿真分析,仿真结果表明：利用本文所设计的控制器,该系统具有较好的动态特性。     

后期模拟制作相机镜头背景模糊效果

在摄影实践中,虚化背景是突出主体的常用手段。但是由于消费级DC镜头的实际焦距都很短,因此实现浅景深而虚化背景的难度较大。如果我们希望用消费级DC也能达到虚化背景突出主体的效果,那么,Photoshop后期处理可以派上用场。     

浅谈随机决策树

该文介绍了随机决策树分类模型及如何启发式选择随机决策树的深度及棵树,通过实验证明了该算法的有效性和高效性。     

具有深度自适应估计的视觉伺服优化

在手眼机器人视觉伺服中,如何确定机器人末端摄像机移动的速度和对物体的深度进行有效的估计还没有较好的解决方法.本文采用一般模型法,通过求解最优化控制问题来设计摄像机的速度,同时,利用物体初始及期望位置的深度估计值,提出了一种自适应估计的算法对物体的深度进行估计,给出了深度变化趋势,实现了基于图像的定位控制.该方法能够使机器人在工作空间范围内从任一初始位置出发到达期望位置,实现了系统的全局渐近稳定且不需要物体的几何模型及深度的精确值.最后给出的仿真实例表明了本方法的有效性.     

灰度／深度图像序列的三维空间内插──一种新的计算机图像生成技术

本文提出的灰度／深度图像序列的三维内插方法,摆脱了原实摄路径限制,可在空间内任意巡行,采用空间稀疏化采样,从而达到大量节省存储量和高度灵活地可控再现空间场景的目的,为解决三维内插中庞大的计算量,我们研究了快速而实用的散点算法和深度适应的Ｗａｒｐ算法。通过介绍三维内插的几种简单情形,导出了一些简单、实用的算法。计算机仿真结果显示了这一新方法是完全可行的。     

Characterization of laser hardened steels by laser induced ultrasonic surface waves

Ultrasonic surface waves are suitable for the characterization of surface hardened materials. This is shown on laser hardened turbine blades. The martensitic microstructure within the surface layer of surface hardened steels has a lower surface wave propagation velocity than the annealed or normalized substrate material. Because the propagation velocity depends on the ratio of layer thickness to wavelengthd/, its measurement allows the determination of the hardening depth. If the surface wave frequency is high enough, the surface wave propagates mainly within the hardened layer. A correlation of the surface wave velocity to the surface hardness has been found. Because the variation of the surface velocity in hardened steels is small, a high measurement accuracy is necessary to obtain the interesting hardening parameters with sufficient certainty. Therefore, a measuring arrangement has been developed where laser pulses, guided by optical fibers to the surface hardened structure, generate simultaneously surface wave pulses at two different positions. The two ultrasonic pulses are received by a piezoelectric transducer. The surface wave velocity is obtained from the time delay between these pulses which is determined by the cross-correlation method. To evaluate simultaneously surface waves with different penetration depths from the same signal acquisition, digital filtering has been used in connection with the cross-correlation.     

基于动态门控特征融合的轻量深度补全算法

稠密深度图在自动驾驶和机器人等领域至关重要,但是现今的深度传感器只能产生稀疏的深度测量,所以有必要对其进行补全.在所有辅助模态中, RGB图像是常用且易得的信息.现今的许多方法都采用RGB和稀疏深度信息结合进行补全.然而它们绝大部分都是利用通道拼接或逐元素求和简单的对两种模态的信息进行融合,没有考虑到不用场景下不同模态特征的置信度.提出一种以输入深度稀疏分布为指导,结合双模态信息量的动态门控融合模块,通过动态产生融合权重的方式对两个模态特征进行更高效的结合.并且根据不同模态的数据特征设计了精简的网络结构.实验结果表明所提出模块和改进的有效性,提出的网络在两个有挑战性的公开数据集KITTI depth completion和NYU depth v2上,使用了很少的参数量达到了先进的结果,取得了性能和速度的优秀平衡.     

基于改进YOLOv3的智慧足球场行人检测

由于足球比赛场景中密集人群、移动小目标居多, YOLOv3算法存在检测精确度较低且模型参数量较大等问题, 使其无法部署在资源算力有限的移动设备上, 本文提出了一种基于改进YOLOv3的行人检测方法, 将Darknet-53主干特征提取网络替换为更加高效且轻量化的GhostNet网络; 同时选取了4个尺度的检测分支层并采用K-means++算法改善anchor box的聚类效果; 添加空间金字塔池化对输入图像实现相同大小的输出; 提出CIoU损失函数来计算目标定位损失值; 添加heatmap热力图可视化并在训练中使用Mosaic数据增强. 实验结果表明, YOLOv3-GhostNet在VOC融合数据集上mAP达到90.97%的同时相比YOLOv3算法提高了1.75%, 参数量减少了约81.4%且实时检测速率提高了约1.5倍, 在小型移动设备上表现出不错的检测效果.     

Travel time models for deep-lane unit-load autonomous vehicle storage and retrieval system (AVS/RS)

Autonomous vehicle storage and retrieval systems use vehicles that move horizontally along rails within the storage racks, while vertical movements are provided by lifts. The solution proposed in this paper addresses a particular system configuration that works with multiple deep storage lanes that are widely used in the food and beverage industry, characterised by large volumes of products of limited variety. The generic deep lane is single item, i.e. one stock keeping unit, and single batch, i.e. one production lot, thereby affecting the performance of the system in terms of storage capacity utilisation and throughput. Determining the number and depth of the lanes is crucial to aid the design and control of such a storage system. The aim of this paper was to support the design of AVS/RSs though a set of original analytic models for the determination of the travelled distance and time for single-command and dual-command cycles given alternative layout configurations. The models are validated by simulation and exemplified with a real-warehousing case study. The paper presents useful guidelines for the configuration of the system layout including the determination of the optimal shape ratio and the length of the lanes.     

基于Kinect传感器的移动机器人环境检测及行为学习

研究了一种基于深度图像和强化学习算法的移动机器人导航行为学习方法。该方法利用机器人装配的Kinect传感器检测工作环境信息,然后对获取的深度图像数据和视频图像进行处理、融合和识别,并由此构建机器人任务学习的状态空间,最终利用强化学习方法实现移动机器人的导航任务的自学习。该方法的有效性通过实验得到验证。实验表明,该方法能够使机器入具有较强的环境感知能力,并能够通过自学习的方式掌握行为能力。