Fast Quaternion Product Units for Learning Disentangled Representations in SO(3)

Real-world 3D structured data like point clouds and skeletons often can be represented as data in a 3D rotation group (denoted as $\mathbb{SO}(3)$). However, most existing neural networks are tailored for the data in the Euclidean space, which makes the 3D rotation data not closed under their algebraic operations and leads to sub-optimal performance in 3D-related learning tasks. To resolve the issues caused by the above mismatching between data and model, we propose a novel non-real neuron model called \textit{quaternion product unit} (QPU) to represent data on 3D rotation groups. The proposed QPU leverages quaternion algebra and the law of the 3D rotation group, representing 3D rotation data as quaternions and merging them via a weighted chain of Hamilton products. We demonstrate that the QPU mathematically maintains the $\mathbb{SO}(3)$ structure of the 3D rotation data during the inference process and disentangles the 3D representations into ``rotation-invariant'' features and ``rotation-equivariant'' features, respectively. Moreover, we design a fast QPU to accelerate the computation of QPU. The fast QPU applies a tree-structured data indexing process, and accordingly, leverages the power of parallel computing, which reduces the computational complexity of QPU in a single thread from $\mathcal{O}(N)$ to $\mathcal {O}(\log N)$. Taking the fast QPU as a basic module, we develop a series of quaternion neural networks (QNNs), including quaternion multi-layer perceptron (QMLP), quaternion message passing (QMP), and so on. In addition, we make the QNNs compatible with conventional real-valued neural networks and applicable for both skeletons and point clouds. Experiments on synthetic and real-world 3D tasks show that the QNNs based on our fast QPUs are superior to state-of-the-art real-valued models, especially in the scenarios requiring the robustness to random rotations.<br>

Fast Quaternion Product Units for Learning Disentangled Representations in SO(3)

Real-world 3D structured data like point clouds and skeletons often can be represented as data in a 3D rotation group (denoted as $\mathbb{SO}(3)$). However, most existing neural networks are tailored for the data in the Euclidean space, which makes the 3D rotation data not closed under their algebraic operations and leads to sub-optimal performance in 3D-related learning tasks. To resolve the issues caused by the above mismatching between data and model, we propose a novel non-real neuron model called \textit{quaternion product unit} (QPU) to represent data on 3D rotation groups. The proposed QPU leverages quaternion algebra and the law of the 3D rotation group, representing 3D rotation data as quaternions and merging them via a weighted chain of Hamilton products. We demonstrate that the QPU mathematically maintains the $\mathbb{SO}(3)$ structure of the 3D rotation data during the inference process and disentangles the 3D representations into ``rotation-invariant'' features and ``rotation-equivariant'' features, respectively. Moreover, we design a fast QPU to accelerate the computation of QPU. The fast QPU applies a tree-structured data indexing process, and accordingly, leverages the power of parallel computing, which reduces the computational complexity of QPU in a single thread from $\mathcal{O}(N)$ to $\mathcal {O}(\log N)$. Taking the fast QPU as a basic module, we develop a series of quaternion neural networks (QNNs), including quaternion multi-layer perceptron (QMLP), quaternion message passing (QMP), and so on. In addition, we make the QNNs compatible with conventional real-valued neural networks and applicable for both skeletons and point clouds. Experiments on synthetic and real-world 3D tasks show that the QNNs based on our fast QPUs are superior to state-of-the-art real-valued models, especially in the scenarios requiring the robustness to random rotations.<br>

Optimization Algorithms for Joint Power and Sub-Channel Allocation for NOMA-Based Maritime Communications

This paper investigates resource optimization schemes in a marine communication scenario based on non-orthogonal multiple access (NOMA). According to the offshore environment of the South China Sea, we first establish a Longley–Rice-based channel model. Then, the weighted achievable rate (WAR) is considered as the optimization objective to weigh the information rate and user fairness effectively. Our work introduces an improved joint power and user allocation scheme (RBPUA) based on a single resource block. Taking RBPUA as a basic module, we propose three joint multi-subchannel power and marine user allocation algorithms. The gradient descent algorithm (GRAD) is used as the reference standard for WAR optimization. The multi-choice knapsack algorithm combined with dynamic programming (MCKP-DP) obtains a WAR optimization result almost equal to that of GRAD. These two NOMA-based solutions are able to improve WAR performance by 7.47% compared with OMA. Due to the high computational complexity of the MCKP-DP, we further propose a DP-based fully polynomial-time approximation algorithm (DP-FPTA). The simulation results show that DP-FPTA can reduce the complexity by 84.3% while achieving an approximate optimized performance of 99.55%. This advantage of realizing the trade-off between performance optimization and complexity meets the requirements of practical low-latency systems.

Design and Implementation of a Wearable Accelerometer-Based Motion/Tilt Sensing Internet of Things Module and Its Application to Bed Fall Prevention

Accelerometer-based motion sensing has been extensively applied to fall detection. However, such applications can only detect fall accidents; therefore, a system that can prevent fall accidents is desirable. Bed falls account for more than half of patient falls and are preceded by a clear warning indicator: the patient attempting to get out of bed. This study designed and implemented an Internet of Things module, namely, Bluetooth low-energy-enabled Accelerometer-based Sensing In a Chip-packaging (BASIC) module, with a tilt-sensing algorithm based on the patented low-complexity COordinate Rotation DIgital Computer (CORDIC)-based algorithm for tilt angle conversions. It is applied for detecting the postural changes (from lying down to sitting up) and to protect individuals at a high risk of bed falls by prompting caregivers to take preventive actions and assist individuals trying to get up. This module demonstrates how motion and tilt sensing can be applied to bed fall prevention. The module can be further miniaturized or integrated into a wearable device and commercialized in smart health-care applications for bed fall prevention in hospitals and homes.

Modular Robotic Limbs for Astronaut Activities Assistance

In order to meet the assist requirements of extravehicular activity (EVA) for astronauts, such as moving outside the international space station (ISS) or performing on-orbit tasks by a single astronaut, this paper proposes an astronaut robotic limbs system (AstroLimbs) for extravehicular activities assistance. This system has two robotic limbs that can be fixed on the backpack of the astronaut. Each limb is composed of several basic module units with identical structure and function, which makes it modularized and reconfigurable. The robotic limbs can work as extra arms of the astronaut to assist them outside the space station cabin. In this paper, the robotic limbs are designed and developed. The reinforcement learning method is introduced to achieve autonomous motion planning capacity for the robot, which makes the robot intelligent enough to assist the astronaut in unstructured environment. In the meantime, the movement of the robot is also planned to make it move smoothly. The structure scene of the ISS for extravehicular activities is modeled in a simulation environment, which verified the effectiveness of the proposed method.

Loop Heat Pipes for Thermal Management of Electric Vehicles

This present paper investigates the potential of loop heat pipe (LHP), with respect to technological merits and application niche, in automotive thermal management. Broadly, LHP design and applicability for hot spot cooling in electronics (local dissipation), and for heat transport over longer distances (remote dissipation) has been proposed and discussed in detail. The basic module in these applications includes loop heat pipe with different shapes and sizing factors. Two types of LHP design have being tested and results discussed. The miniature version, with 10 mm thick and flat evaporator, for cooling ECU with 70 W chipset while keeping source temperature below 100 °C limit was evaluated. Two larger versions with cylindrical evaporator, 25 mm diameter & 150 mm length, and heat transfer distances of 250 mm and 1000 mm respectively were tested for power electronics and battery cooling, with more than 500 W transport capabilities in gravity field. In conclusions, loop heat pipes will provide an energy efficient passive thermal control solution for next generation low emission automotive, particularly for electric vehicles which have high level electrifications and more definitive cooling requirements.

Student Behavior Recognition From Heterogeneous View Perception in Class Based on 3-D Multiscale Residual Dense Network for the Analysis of Case Teaching

The study of student behavior analysis in class plays a key role in teaching and educational reforms that can help the university to find an effective way to improve students' learning efficiency and innovation ability. It is also one of the effective ways to cultivate innovative talents. The traditional behavior recognition methods have many disadvantages, such as poor robustness and low efficiency. From a heterogeneous view perception point of view, it introduces the students' behavior recognition. Therefore, we propose a 3-D multiscale residual dense network from heterogeneous view perception for analysis of student behavior recognition in class. First, the proposed method adopts 3-D multiscale residual dense blocks as the basic module of the network, and the module extracts the hierarchical features of students' behavior through the densely connected convolutional layer. Second, the local dense feature of student behavior is to learn adaptively. Third, the residual connection module is used to improve the training efficiency. Finally, experimental results show that the proposed algorithm has good robustness and transfer learning ability compared with the state-of-the-art behavior recognition algorithms, and it can effectively handle multiple video behavior recognition tasks. The design of an intelligent human behavior recognition algorithm has great practical significance to analyze the learning and teaching of students in the class.

Modular Timber Gridshells

Needs of considering environmental impacts and growing emphasise on using renewable resources is discussed also in the field of architecture and construction. Timber, as a key element, provides certain benefits with less negative environmental impact than non-renewable resources, represented by commonly used materials (e.g. concrete). Nevertheless, using the timber as a core element for constructing of multi-storey or large hall buildings can be more difficult due to limits of the wood. New task for architects and engineers therefore arises and the target is to come up with a feasible solution for using the timber as a core material, despite its features. One of the key requirements for wide use of load bearing timber structure systems is “modularity”. Many different forms of structures can be achieved by repeating the basic module to fulfil architect’s intentions. This article focuses on a large-span structure, based on gridshell type systems. Such modules could be simply prefabricated and connected to another module. Modules can be pretty small, which contributes to easy and cost-effective transportation. Nevertheless, technical requirement and limitation of such modules needs to be taken into consideration. Problems come up with using the same universal element in parts of structure with different character of load and stress. Also, universal element should fulfil various space and form requirements set by architect. Suggested construction system is based on “fractality”. The goal of the system is to fulfil requirement of final structure and space requirements by replacing the standard module by a predetermined number of sub-modules. Structure is then denser in exposed areas and sparser in others. Design module is than expected to be tested and put into practice for wide usage in construction projects where wood is desired as the core material.

Analysis of Tai Chi Ideological and Political Course in University Based on Big Data and Graph Neural Networks

Tai Chi martial arts education is one of the components of school education. Its educational value is not only to require students to master basic Tai movement technical skills and improve their physical fitness but also to bring students’ ideological progress and cultivate students to respect teachers and lectures. Excellent moral qualities such as politeness, keeping promises, observing the rules, and acting bravely, as well as the cultivation of the spirit of unity and cooperation, and the quality of will also have a certain meaning. However, the scientific Tai Chi ideological and political courses and the construction of Wude education interactive classrooms lack relevant research. Therefore, this article builds a Tai Chi ideological and political interactive classroom system based on big data technology and graph neural network. First, the spatio-temporal graph convolutional neural network is used to reason about the relationship between Tai Chi action categories and strengthen the low-dimensional features of semantic categories and their co-occurrence expressions used for semantic enhancement of current image features. In addition, in order to ensure the efficiency of the Tai Chi scene analysis network, an efficient dual feature extraction basic module is proposed to construct the backbone network, reducing the number of parameters of the entire network and the computational complexity. Experiments show that this method can obtain approximate results, while reducing the amount of floating-point operations by 42.5% and the amount of parameters by 50.2% compared with the work of the same period, and achieves a better balance of efficiency and performance. Secondly, based on the big data of historical Tai Chi classrooms, this article constructs an interactive classroom system that can effectively improve the quality of Tai Chi ideological and political courses.

Unstructured Road Segmentation Based on Road Boundary Enhancement Point-Cylinder Network Using LiDAR Sensor

The segmentation of unstructured roads, a key technology in self-driving technology, remains a challenging problem. At present, most unstructured road segmentation algorithms are based on cameras or use LiDAR for projection, which has considerable limitations that the camera will fail at night, and the projection method will lose one-dimensional information. Therefore, this paper proposes a road boundary enhancement Point-Cylinder Network, called BE-PCFCN, which uses Point-Cylinder in order to extract point cloud features directly and integrates the road enhancement module to achieve accurate unstructured road segmentation. Firstly, we use the improved RANSAC-Boundary algorithm to calculate the rough road boundary point set, training in the same parameters with the original point cloud as a submodule. The whole network adopts the encoder and decoder structure, using Point-Cylinder as the basic module, while considering the data locality and the algorithm complexity. Subsequently, we made an unstructured road data set for training and compared it with existing LiDAR(Light Detection And Ranging) semantic segmentation algorithms. Finally, the experiment verified the robustness of BE-PCFCN. The road intersection-over-union (IoU) was increased by 4% when compared with the best existing algorithm, reaching 95.6%. Even on unstructured roads with an extremely irregular shape, BE-PCFCN also currently has the best segmentation results.