The head shape of high-speed trains has become a critical factor in boosting the speed further. Aerodynamic simulation-based optimization is a dominant method to obtain the optimal head shape which relies on detailed train head models defined by a lot of design variables. Since aerodynamic simulation-based optimization involves heavy calculations, too many design variables not only causes high computational costs, but also makes the optimal solution difficult to obtain. Therefore, how to use few design variables to define detailed train head model is the key to success. Partial differential equation (PDE)-based geometric modelling which creates a complicated PDE patch with few design variables provides an effective solution to this problem. In addition, it also has the advantage of naturally maintaining any high-order continuities between two adjacent surfaces which is very important in designing highly smooth train heads to achieve excellent aerodynamic performance. At the present time, PDE-based geometric modelling cannot be directly applied in computer-aided design (CAD), computer-aided manufacturing (CAM), and computer-aided engineering (CAE) since it has not become an industrial standard. In contrast, non-uniform rational B-splines (NURBS) are commonly used in CAD, CAM, CAE, and many other engineering fields. They have already become part of industry wide standards. In order to apply PDE-based geometric modelling in shape design of high-speed train heads for CAD etc., how to optimally convert PDE surfaces into NURBS surfaces must be addressed. In this paper, a new method of achieving optimal conversion of PDE surfaces representing high-speed train heads into NURBS surfaces is developed. It takes control points and weight deformations of NURBS surfaces to be design variables, and the error between NURBS surfaces and PDE surfaces as the objective function. The least squares fitting and the genetic algorithm are combined to obtain the optimal conversion between PDE surfaces and NURBS surfaces. The application examples demonstrate the effectiveness of the developed method.
11 October 2023 in IEEE Transactions on Multimedia
Exploiting the rich temporal information in human pose sequences to facilitate 3D pose estimation has garnered particular attention. While various learning architectures have been designed for temporal exploiting, these architectures are usually trained via the 3D pose loss independently imposed on every single frame, without explicit temporal signals introduced for supervision. This inevitably increases the difficulty of temporal exploiting, since the network must reason about the meaningful temporal information based on the non-temporal single-frame supervision first. Only then, the network can utilize this information to guide sequence modeling. Recently, some work introduce temporal smoothness as an explicit supervision signal, which makes the network more straightforwardly reaches the temporal information from the supervision signal, thus improving the temporal exploiting. However, the temporal smoothness only roughly measures the short-term temporal properties between adjacent frame pairs. In this work, we propose to generalize the supervision of temporal smoothness to temporal correlations, letting the network precisely consider more comprehensive temporal properties in sequences. We contribute two novel correlation-based loss functions, which adopt different strategies to respectively regularize the encoder and decoder sides of the network for temporal exploiting. Besides, we design a pre-training scheme to ensure a general convergence of existing pose estimators under our correlation losses. Experiments on three benchmarks demonstrate that our method can be compatible with different networks, improving their temporal exploiting ability to output more accurate and robust pose estimations.