基于深度高斯过程回归的视频异常事件检测方法

针对现有异常检测方法忽视异常事件发生概率小而造成虚警这个问题,基于高斯过程回归(GPR)的框架,将GPR核函数非参数化所具有的灵活性与深度神经网络的结构特性相结合,并将卷积神经网络封装在GPR的核函数中,以同时实现异常检测任务中特征提取和检测两个步骤。在测试阶段,相对于训练样本集的后验概率的对数似然较小的被判定为异常。方法在一个模拟数据集和一个完全真实的数据集上进行了实验验证,实验结果证明所提出的方法在两个数据集上分别达到了83.9%的帧级AUC和34.4%的帧级AUC,在性能上达到了现有技术发展水平。     

Effect of pre-annealing on thermal stability of amorphous Zr-Cu-Ni alloy

The influence of pre-annealing on thermal stability of the amorphous Zr70Cu20Ni10 alloy was reported by em-ploying the differential scanning calorimetry (DSC) and high-resolution transmission electron microscopy (HRTEM) tech-niques. It has been observed that the supercooled liquid region decreases with increasing the annealing time under isothe-mal conditions, indicating that the thermal stability of the amorphous Zr70Cu20Ni10 alloy decreases gradually. HRTEM ob-servations reveal that there exist some ordered atomic clusters in the amorphous matrix at the relaxation stage. These or-dered atomic clusters can be regarded as precursors for the precipitation of the crystalline phases in the subsequent crystal-lization process. The reasons resulting in the decrease in thermal stability of the amorphous Zr70Cu20Ni10 alloy with anneal-ing time are discussed through the Gaussian decomposition for the radial distribution function of the amorphous Zr70Cu20Ni10 alloy.     

Optimisation of hybrid tandem metal active gas welding using Gaussian process regression

In this paper, an additional filler wire with opposite polarity was inserted in tandem flux cored arc welding process to increase the welding speed and deposition rate. In this hybrid welding, the optimisation of welding parameters is required to improve the bead geometry which directly indicates the welding quality. However, the correlation between the parameters and the bead geometry is hard to identify, so the process parameters are usually selected intuitively by the experienced engineers. Therefore, welding process modelling is constructed with the Gaussian process regression model, and parameter optimisation is performed with sequential quadratic programming optimisation algorithm. The proposed modelling optimisation process is verified by performing the welding experiment using the parameters that are optimised by the proposed process.     

Finite element prediction of material removal rate due to electro-chemical spark machining

Electro-chemical spark machining (ECSM) is an innovative hybrid machining process, which combines the features of the electro-chemical machining (ECM) and electrodischarge machining (EDM). Unlike ECM and EDM, ECSM is capable of machining electrically non-conducting materials. This paper attempts to develop a thermal model for the calculation of material removal rate (MRR) during ECSM. First, temperature distribution within zone of influence of single spark is obtained with the application of finite element method (FEM). The nodal temperatures are further post processed for estimating MRR. The developed FEM based thermal model is found to be in the range of accuracy with the experimental results. Further the parametric studies are carried out for different parameters like electrolyte concentration, duty factor and energy partition. The increase in MRR is found to increase with increase in electrolyte concentration due to ECSM of soda lime glass workpiece material. Also, the change in the value of MRR for soda lime glass with concentration is found to be more than that of alumina. MRR is found to increase with increase in duty factor and energy partition for both soda lime glass and alumina workpiece material.     

基于Surfacelet变换和动态纹理的烟雾检测

鉴于烟雾检测对火灾预警的重要作用,提出一种基于Surfacelet变换的动态纹理烟雾检测算法。先对图像序列进行Surfacelet变换,再对变换后的系数进行广义高斯建模,获得与系数相对应的模型参数作为特征,最后使用KL距离做相似性度量。与其他3种基于Surfacelet变换的烟雾检测方法进行对比,包括：使用均值和方差作为特征,支持向量机进行分类;使用均值和方差作为特征,欧式距离进行相似性度量;使用广义高斯模型参数作为特征,欧式距离进行相似性度量。实验结果表明,该算法可以提高烟雾检测准确性,降低误检率,有效去除类烟运动物体的干扰。     

一种基于高斯混合模型的轨迹预测算法

在智能交通控制系统、军事数字化战场、辅助驾驶系统中,实时、精确、可靠的移动对象不确定性轨迹预测具有极高的应用价值.智能轨迹预测不仅可以提供精准的基于位置的服务,而且可以提前监测和预判交通状况,进而推荐最佳路线,已经成为移动对象数据库研究的热点,亟需设计准确而高效的位置预测方法.针对现有方法的不足,提出了基于高斯混合模型的轨迹预测方法GMTP,主要步骤包括:(1) 针对复杂运动模式利用高斯混合模型建模;(2) 利用高斯混合模型计算不同运动模式的概率分布,进而将轨迹数据划分为不同分量;(3) 利用高斯过程回归预测移动对象最可能的运动轨迹.GMTP是高斯非线性概率统计模型,其优势在于:计算结果不仅是位置预测值,更是关于移动对象未来所有可能运动轨迹的概率分布,可以利用概率统计分布特性获得某种运动模式(如匀加速运动)下的位置预测.大量真实轨迹数据集上的实验结果表明:与相同参数设置下的高斯回归预测和卡尔曼滤波预测法相比,GMTP的预测准确性平均提高了22.2%和23.8%,预测时间平均缩减了92.7%和95.9%.     

基于多尺度核函数的铆接件腐蚀疲劳预测

目前腐蚀疲劳破坏预测方法精度不高.提出基于小波多分辨分析法(MRA),在再生核希尔伯特空间构建一种多尺度核函数的最小二乘支持向量机(multi-scale kernel LSSVM,MSK_LSSVM)预测算法.根据Mercer平移不变核定理,构造了多尺度复Gaussian小波核函数.由于多尺度核函数能够通过平移生成L2(R2)子空间的一组完备基,因此MSK_LSSVM可以任意逼近目标函数,更具灵活性.经仿真实验验证,与BP神经网络方法、标准支持向量机、灰色系统预测模型方法对比,机械结构中铆接件腐蚀变化的趋势通过MSK_LSSVM预测,准确率高、时间短.     

一种基于像素梯度信息的背景减除法

讨论了背景模型的更新参数与模型精度的关系。通过精确的梯度背景模型值间接估计当前帧中背景像素理论上的期望梯度值。以高斯模型为基础,将当前帧背景像素的实际梯度值与其理论上的期望值进行比较,计算偏差概率,以此为基础,形成不依赖于局部纹理的梯度特征的相似性度量方法。再用梯度特征的相似度量化地调整差分图像在各像素点处的二值化阈值,实现像素值信息与梯度信息的融合使用。实验表明,本方法对前景分割有一定的改善效果。