Co-speech gesture generation, a subset of 3D motion generation, aims to generate appropriate motion sequences from audio or other conditions. While many existing methods focus on the rhythm between motion and audio, they often neglect the semantics of gestures. Furthermore, approaches based on diffusion models or Transformer require significant time for training and inference, making them unsuitable for real-time applications. In this paper, we proposed CASG, a network based on the conditional autoregressive model, which effectively enhances the semantics of generated gestures through a semantic enhancement module inspired by VQ-VAE. Additionally, the loss function is improved for 3D rotational and translational transformations of motion sequences, addressing the instability issue in generated models. Extensive experiments demonstrate that our model outperforms competing methods in terms of semantics, rhythm and stability.

To address the issue of edge thickness affecting surface quality and color in 3D printing, particularly when dealing with variations in the thickness of upper and lower surfaces, a surface isometric offset transformation method is employed to obtain the optimal subsurface structure. Based on these offset surfaces, adaptive offset calculations are performed for the exterior surface structure using different offset distances for different surfaces. Subsequently, each surface is iteratively optimized, redesigning the surface structure of the printed part to meet the requirements for surface quality and color depth. The results show that the adaptive offset method for surfaces is superior to the traditional edge line offset method in 3D printing. It provides an effective surface treatment method for color printing, especially under requirements for translucency and light colors, allowing for customized surface thickness settings based on parameters such as surface hardness and brightness, thereby enabling adaptive surface thickness.

Convolutional Neural Networks (CNNs) have demonstrated significant advantages in binocular stereo matching tasks, but the high computational complexity of their 3D extensions (3D CNNs) limits real-time applicability. This study proposes a 2D convolutional cost feature-based binocular matching neural network accelerator, achieving efficient deployment through algorithm-hardware co-design. Key innovations include: (1) Designing the FDSCS network architecture with optimized cost volume generation modules, integrating enhanced preprocessing and pipeline mechanisms; (2) Restructuring 2D convolution via an Img2Col-GEMM strategy to leverage FPGA parallel computing for accelerated matrix operations; (3) Introducing network weight quantization and bilinear interpolation modules to reduce memory requirements while improving output accuracy. Evaluations on the ZCU102 platform demonstrate that the accelerator achieves 18.72 ms per-frame processing speed and a 3.22% average error rate on the KITTI dataset, balancing real-time performance with precision for time-sensitive applications such as autonomous driving.

The application of Convolutional Neural Networks (CNNs) has significantly accelerated the development of semantic segmentation, particularly in the domain of autonomous driving. Semantic segmentation is crucial for enabling autonomous vehicles to accurately perceive their surroundings and make real-time decisions. However, the increasing computational complexity has led to a substantial rise in power consumption, hindering the progress of self-driving technology. While Field Programmable Gate Arrays (FPGAs) offer a means to accelerate network inference, achieving an optimal balance between accuracy and speed remains a significant challenge. This paper investigates state-of-the-art semantic segmentation models and their corresponding optimization techniques. We summarize the critical requirements for system flexibility when mapping models to embedded FPGAs. Based on these requirements, we propose a reconfigurable semantic segmentation accelerator that integrates hardware optimization and data quantization strategies. The data quantization strategy reduces the bit width to 8 bits without any discernible loss in accuracy. To further reduce inference time, the network operators are implemented and optimized directly in hardware. Additionally, an instruction-controlled data flow is employed to support future updates and scalability. To enhance coding efficiency and reusability, we utilize SpinalHDL, an emerging hardware description language embedded in Scala, a high-level programming language, for the development of the proposed accelerator. The performance of the design is evaluated on the Virtex UltraScale+ VU9P FPGA platform, yielding accuracies of 74.3% on the CamVid dataset and 72.1% on the Cityscapes dataset, with a processing speed of 24 FPS, approaching real-time performance. This work paves the way for more energy-efficient and scalable solutions for autonomous driving systems, with potential for real-world deployment in various safety-critical environments.

Under the condition of double-track parallel operation, the subgrade structure is subjected to the superimposed effects of intersecting train loads, resulting in complex dynamic response characteristics. Taking the subgrade engineering at the entrance of Limin Tunnel on the Harbin-Mudanjiang Passenger Dedicated Line as the research background, this study established a finite element model of the CRTS III slab ballastless track-subgrade-foundation structure using the ANSYS APDL module. Combining multi-body dynamics methods, the indirect coupling method was employed to analyze the dynamic responses of both the subgrade and track slab structures under train loading. The results indicate that stress concentration occurs at the interface between the subgrade and track slab. The vertical displacement time-history curves at different subgrade positions exhibit an initial increase followed by a decreasing trend, with vertical dynamic displacements forming a semi-elliptical distribution pattern under the track structure. The vibration-induced dynamic displacements in the subgrade attenuate along the depth direction and superimpose near the track centerline, reaching 1.26 mm vertical dynamic displacement at the subgrade base surface accompanied by 0.17 mm residual deformation. Particular attention should be paid to the soil strength at track slab edges and the central subgrade area. This research provides theoretical support for the design optimization and long-term service performance evaluation of double-track ballastless track subgrades.

This paper investigates the influence of technological parameters such as cutting voltage (corresponding to cutting height) and cutting speed on the surface quality of the cut when machining with a CNC plasma cutting machine. The experiment was conducted using an HNC 1500W CNC plasma cutting machine on SS400 steel with a thickness of 5mm. Surface quality was evaluated based on the surface roughness of the cut. The Taguchi experimental design method was used to assess the impact of each individual parameter on surface roughness. The minimum surface roughness achieved was 1.5 μm, corresponding to a cutting voltage of 120 V and a cutting speed of 1700 mm/min. The analysis results using the Taguchi method indicate that surface roughness is mainly influenced by cutting voltage (which corresponds to the cutting height), while cutting speed has a lesser effect.

This paper presents a wireless pressure signal real-time acquisition and waveform display system based on embedded technology. The system takes the STM32F103RCT6 microcontroller as the core, and combines the high-precision ADC analog-to-digital conversion module, LCD display module and serial communication module to realize the digital acquisition, processing and visualization of pressure signals. Through the synchronous working mode of ADC+DMA+TIMER, the system can dynamically adjust the sampling frequency and store the data in real time, and use the FFT algorithm to complete the frequency domain analysis, and at the same time, realize the parameter adjustment through external interrupt. The experimental results show that the system can accurately collect the pressure signal and display the waveform in real time through the LCD screen, and the serial communication function supports the data analysis of the host computer. The test data show that the average value of sensor output voltage is negatively correlated with the pressure in the pressure range of 50g to 200g weights (R²>0.98), which verifies the reliability and accuracy of the system. The design provides an efficient solution for monitoring pressure signals in industrial monitoring, medical equipment and other fields.

In the field of target detection and image classification, although hardware accelerators for CNN and Transformer are widely available, accelerators specifically designed for Transformer-CNN hybrid networks are relatively rare. Due to the high computational complexity, storage requirements, memory bandwidth limitations, and parallel computation difficulties of hybrid networks, the implementation of the model on hardware is challenging. To address these issues, this study proposes a dedicated hardware accelerator with a configurable Systolic Array computational architecture, on the one hand, designing an Img2col module for converting 3D feature maps into 2D matrices, and utilising the Systolic Array to implement convolution operations with different sizes as well as matrix multiplications, which is specifically designed to accelerate the inference of Transformer-CNN hybrid networks. On the other hand the accelerator is extremely flexible and configurable, parametrically configuring the Systolic Array according to computational needs ensures that its performance is fully exploited, while the use of on-chip buffers for storing maps, weight data, and intermediate results reduces off-chip memory accesses and power consumption, and improves data reusability.

Our well-designed accelerator was tested on Xilinx Zynq, and the experimental results show that the accelerator exhibits excellent performance in both CNN and Attention Mechanism computation, with an output of up to 608.6 GOPS/W. The aim of this study is to build efficient hardware accelerators that utilise efficient computational units and memory structures for Transformer-CNN hybrid network to accelerate the processing and improve the performance of CNN and Transformer deployed on hardware.

In the context of bottom plate inspections for oil tanks, tethered unmanned vehicles (SLUVs) encounter cable entanglement issues due to the complex environment created by multiple support columns within the tanks. This paper introduces an innovative cable anti-entanglement algorithm based on adaptive path planning, aiming to optimize the inspection path of SLUVs in multi-column oil tank environments to achieve efficient and comprehensive bottom plate inspections. The core idea of the algorithm is to intelligently plan the SLUV's motion trajectory to avoid potential entanglement risks with support columns while ensuring comprehensive scanning of the tank's bottom plate. We conducted a series of simulations and real-world tests to verify the algorithm's effectiveness. The results show that the algorithm significantly improves inspection efficiency and prevents work interruptions due to entanglement, thereby greatly reducing the need for manual intervention. This study not only offers an efficient solution for oil tank bottom plate inspections but also opens up new possibilities for tethered robots in similar multi-column environments. With further research and development, the technology's application range is expected to expand to more industrial and exploratory fields, providing robust technical support for automated operations in complex environments.

Transformer models based on attention mechanisms have shown superior performance in the field of computer vision. Designing dedicated accelerators for Transformers can significantly enhance inference performance and reduce power consumption. In this paper, we first employ an integer quantization strategy for both features and weights to reduce the storage requirements and computational complexity of the model. Then, based on the computational characteristics of convolution and attention mechanisms, we design an efficient and flexible hardware architecture, including a multidimensional systolic array and nonlinear normalization acceleration units. This architecture efficiently maps algorithms to the systolic array, optimizing data storage efficiency and minimizing data movement. Additionally, the accelerator design is implemented on an FPGA. Experimental results show that the proposed FPGA accelerator improves model inference speed with minimal accuracy loss. The ViT-base model achieves an average throughput of 626.2 GOPS on this accelerator, with a synthesized power consumption of 16.4W at a 200MHz clock frequency.

Due to the significant reduction in visibility during foggy weather, it poses a considerable challenge for autonomous driving object detection in foggy conditions. There are numerous existing methods to enhance detection performance; however, they do not simultaneously address the issues of detection accuracy, detection speed, and the scarcity of labeled datasets.This paper proposes a rapid detection algorithm based on YOLOv10 and unsupervised domain adaptation for autonomous driving object detection in foggy weather. Firstly, to tackle the issue of limited labeled datasets, the concept of unsupervised domain adaptation is employed, utilizing adversarial learning to achieve domain alignment between foggy and clear weather conditions. This allows for training with only labeled clear weather images and unlabeled foggy images. Secondly, as most current achievements focus solely on detection performance, the paper addresses the speed issue by adopting YOLOv10 without non-maximum suppression (NMS) post-processing as the fundamental detection framework, significantly improving inference speed. Lastly, due to the anchor-free and NMS-free nature of the network, the generalization performance may decline, especially for small-sized targets. The paper replaces the CIOU-based box loss with WNWDLoss to enable the network to handle targets of various sizes, thereby enhancing detection accuracy. Experimental results on the public datasets Cityscapes, FoggyCityscapes and KITTI demonstrate the effectiveness of the proposed method.

Small watershed rivers, due to their limited pollution capacity and weak self-purification function, are easily affected by domestic sewage and agricultural non-point source pollution, leading to water body blackening and ecological degradation. Traditional treatment technologies rely on interception and chemical treatment, which come with high costs and significant risks of ecological damage. This study proposes a water quality improvement technology based on the concept of ecological restoration, integrating biological film aeration, microbial enhancement, and the synergistic effect of aquatic plants. Through practical river engineering projects in Nan 'an City (such as Pengmei Creek and Liankeng Creek), this technology has been verified to significantly increase dissolved oxygen (DO ≥ 2 mg/L), reduce pollutants (COD removal rate ≥ 75%, ammonia nitrogen removal rate ≥ 80%), and restore the water body's self-purification ability. The research results provide an economically efficient and environmentally friendly solution for small watershed river management, with significant theoretical and practical value.