空地融合的城市道路基础设施数字化技术研究

城市道路基础设施三维模型的重构,在城市道路BIM应用与数字化领域具有重大意义;针对城市交通基础设施数字化重构的需求,对车载激光扫描与无人机倾斜摄影采集技术进行综合运用,提出一套从信息采集、空地点云配准、点云分割到三维重构的完整技术方法;首先使用车载激光扫描技术和无人机倾斜摄影技术对交通基础设施信息进行采集,并使用运动恢复结构算法(SfM,structure from motion)生成基础设施空地点云;其次使用迭代最近点法(ICP,iterative closest point)对空地点云进行精配准,然后利用基于PointNet网络的方法对融合点云进行语义分割;最后对分割出的交通基础设施对象进行三维重构;提出的空地融合的城市交通基础设施数字化技术能够高效地实现交通基础设施重构,为城市交通基础设施数字化提供基础、为后续交通专业领域的应用研究提供便利.     

基于微服务的深度学习模型服务系统

Tensorflow Serving是Google开源的一个服务系统,针对Tensorflow Serving单体应用吞吐量低、服务调用烦琐、模型生命周期管理不完善等问题,本文设计了一种基于Tensorflow Serving的微服务软件架构方案,在部署Tensorflow Serving的Docker(开源的应用容器引擎)容器里添加本文研发的监控程序,该监控程序根据各个实例模型加载情况,将可用模型服务主动注册到微服务架构中的注册中心以实现对模型的编排管理。实验结果表明:采用本文的微服务架构方案,有效提升了Tensorflow Serving服务的吞吐量,降低了服务响应时间,简化了模型调用流程,从而满足Tensorflow Serving在生产环境中部署和运维的实际需求。     

Robust registration based compensation method based on sparse recovery

The loss of the degree of system freedom and mismatch of prior information based on the registration based compensation (RBC) method will result in a decline in computational performance. In order to solve above problems, a robust method named SRBC is proposed where the sparse recovery method is applied instead of the sub-aperture smoothing operation. First, the SRBC utilizes the sparse recovery to get the super resolution clutter space-time spectrum. Then the transition covariance matrix is calculated. Finally, the clutter covariance matrix is reconstructed by the Capon spectrum. Compared with the traditional RBC method, the SRBC proposed does not depend on sub-aperture smoothing operation and prior information and maintains the performance of clutter range-dependence. In addition, it is also steadier when the sensor error exists. Experimental simulation demonstrates the correctness and feasibility of this method.     

Registration based compensation method based on adaptive subspace tracking

The registration based compensation (RBC) method is an effective method to compensate the range-dependence of the main-lobe clutter and side-lobe clutter in the same time. However, the compensation performance of the RBC degrades because of the mismatch of prior information and the loss of degree of system freedom. Moreover, the RBC is not very suited for real-time implementation because of the enormous computational complexity and memory usage of eigenvalue decomposition. In this paper, a novel clutter range-dependence compensation method using the modified maximum likelihood adaptive subspace estimation algorithm, which is named the MRBC method for short, is proposed. The eigenvectors matrix and eigenvalues matrix of the clutter covariance matrix are estimated by iterative tracking instead of temporal and spatial smoothing, spectrum calculation and eigenvalue decomposition. Compared with the traditional RBC method, the proposed method can reduce the computational complexity significantly and maintain the performance of clutter range-dependence compensation. In addition, the proposed method can also achieve good performance when the system error exists because of no use of prior information. Experimental simulations demonstrate the validity of this method.     

Adaptive particle swarm optimization method with stagnancy information

To improve the performance of the particle swarm optimization algorithm, the optimal network of the particle age structure with stagnation information is designed, and the information about this network is used to adaptively change the three key parameters of the particle swarm optimization algorithm. At the same time, an adaptive particle swarm optimization method with stagnancy information is proposed and specific optimization steps of this method are given. Four classical low and high dimension benchmark test functions are used to validate the performance of the optimization method, and a comparison study is made with gravitational search algorithm and the traditional particle swarm optimization algorithm without stagnancy information. The comparison study shows that the search efficiency of the proposed method is 2 times higher than that of other methods in the literature in the case of low dimensional multimodal functions. When the dimension of functions is higher than 2, the search efficiency of the proposed method is almost the same as that of other methods, but with the better ability to achieve global solution and local solutions, and the higher solving precision.     

一种鲁棒的多态人脸识别算法

如何处理人脸识别中的多态性一直是人脸识别领域的一个难题。传统的图嵌入算法忽视了同类中多态子类间的同属关系,而且也没有恰当地处理异类间的区别信息。该文提出一种鲁棒的图嵌入人脸识别算法,该算法可以恰当地模拟同类中的多态间关系,而且能在局部流形结构与全局区别信息间实现平衡。基于多个公开数据库的人脸识别实验证明了该方法的有效性。     

基于模拟退火算法的三维人脸匹配方法

基于模拟退火算法,利用深度信息对三维人脸进行匹配和识别。通过具有全局优化能力的模拟退火算法搜索获取全局极值,选择合适的适应度函数MLESAC和SIM控制匹配过程获得识别结果。基于“由粗到细”的思想,分3步匹配三维人脸,并对精准匹配阶段进行改进,通过选择合适的区域和分类器能够更好地应对表情的影响。实验结果表明,模拟退火算法跳出了局部最优解而获得了全局优化解,同时通过选择MLESAC和SIM作为不同匹配阶段的适应度函数能够有效控制匹配过程,提高了识别精准率。     

水下结构物修复用先进机械手软硬件关键技术综述

机器人进行水下结构物维修时,不仅处于非结构化环境,而且存在浮游碰撞危险,作业区域可达性差,水下环境识别困难等特殊问题,所以,进一步提升机器人智能化水平非常重要。英国GRL公司用于焊缝检验的先进海底机械手ARM系统采用先进的人机接口以及工厂用机器人的某些技术,实现了机械手的控制。以ARM的开发经验为基础,目前GRL公司正在研制更为先进的机器人控制系统Polecat,该系统将用于ROV导航员的训练平台Rovsim以及一系列欧盟和商业项目之中,其适用场合不仅包括海洋工程,也包括核电和其他环境。     

一种基于局部窗口的SAR图像目标检测算法

针对合成孔径雷达(SAR)目标检测精确性、实时性和鲁棒性的要求,设计了一种基于局部窗口的SAR图像目标检测算法。该算法在对获取的SAR图像进行去噪和分割处理的基础上,基于尺度不变特征变换(SIFT)实现了亚像素精度快速配准策略;同时,通过SIFT特征的描述结果降维和基于局部窗口的最大期望算法(EM)实现了目标检测。实验结果表明,该算法对复杂背景和光照、旋转变化有较强的自适应性,获得了理想的目标检测效果。     

2D registration based on contour matching for partial matching images

The mean Hausdorff distance, though highly applicable in image registration, does not work well on partial matching images. An improvement upon traditional Hausdorff-distance-based image registration method is proposed, which consists of the following two aspects. One is to estimate transformation parameters between two images from the distributions of geometric property differences instead of establishing explicit feature correspondences. This procedure is treated as the pre-registration. The other aspect is that mean Hausdorff distance computation is replaced with the analysis of the second difference of generalized Hausdorff distance so as to eliminate the redundant points. Experimental results show that our registration method outperforms the method based on mean Hausdorff distance. The registration errors are noticeably reduced in the partial matching images.