Precise perception of road conditions and long-distance traffic signs constitutes the core prerequisite for reliable decision-making in autonomous driving systems, while complex background interference remains a critical bottleneck for such detection tasks. To address this issue, this paper proposes a lightweight detection model (CS-DETR) based on Real-Time Detection Transformer (RT-DETR). Specifically, we first design a channel-separable gated progressive deformable convolution (CS-GPDC) that combines lightweight characteristics with receptive field enhancement capabilities. Second, we construct a lightweight backbone network (CSLIT) centered on this convolution, integrating multi-head attention (MHA) and ConvBN techniques. Finally, we introduce a resolution-aware position fusion module (RAP-FM), which enables efficient cross-scale feature interaction and fusion through adaptive weight adjustment, thereby further enhancing feature representation capabilities. Experimental results demonstrate that the CS-DETR model achieves a mean average precision (mAP) of 96.1% on the KITTI dataset, representing a 4.4 percentage point improvement over the RT-DETR model. On the GTSDB dataset, it attains an mAP of 97.3%, delivering a 2.6 percentage point enhancement relative to RT-DETR. Meanwhile, the model reduces its parameter count by 12.9 million, achieving a dual optimization of detection accuracy and lightweight performance while ensuring real-time detection capabilities. Ablation experiments validate the effectiveness of the proposed feature extraction and cross-scale fusion modules in boosting detection accuracy.
Multimodal 3D understanding has rapidly emerged as a critical research area, driven by applications in autonomous navigation, robotics, and augmented reality. Following PRISMA guidelines, we systematically searched four databases and identified 53 studies integrating multimodal large language models (MLLMs) with 3D point cloud data, published between 2023 and 2026. We propose a two-level taxonomy distinguishing 3D scene understanding (indoor, multi-object) from 3D object understanding (single-object, fine-grained), since this distinction shapes both training methodology and applicable tasks. For each level, we review task types, benchmarks, training strategies, feature extraction, and alignment mechanisms, cross-comparing reported results within each task family. Our analysis finds that purpose-built spatial and alignment modules, and training data scale and quality, are stronger predictors of performance than model or parameter size. We further identify evaluation reliability as a systemic, largely unresolved bottleneck: automatic metrics are biased toward short or superficially similar outputs, and LLM-based judges introduce instability, cost, and reproducibility concerns of their own. We conclude with a ranked set of open challenges and recommendations for future research, spanning near-term evaluation fixes, medium-term architectural and data priorities, and long-term infrastructure needs. This survey aims to serve as a foundational reference for those exploring the evolving field of multimodal 3D understanding.
Over the last decade, we have witnessed a large interest of the scientific community in developing innovative techniques based on artificial vision for the automated detection and counting of individuals in images and videos. Despite advancements, counting people in highly crowded environments remains an unresolved challenge. In this paper, we propose a real-time algorithm for crowd counting based on point detection, able to handle both sparse and very dense situations, even in the presence of limited training data for high-density scenarios, by a combining curriculum learning strategy with a semi-supervised approach. In fact, during the training procedure, samples are presented to the network starting with simpler concepts (i.e., images with a few people) and gradually advancing to more complex ones (i.e., images with high density). Additionally, to address the challenge of limited labeled data in very crowded environments, the semi-supervised component is selectively activated only in the latest curriculum stages, namely the medium-hard and hard density levels, where unlabeled data become particularly valuable. The experiment, conducted on a test set composed of about 44,000 images belonging to 18 different publicly available datasets, confirms the effectiveness of the proposed approach. Compared with a state-of-the-art baseline point detection algorithm, we observe impressive average improvements of 73% and 70% in terms of Mean Absolute Error (MAE) and Mean Squared Error (MSE), respectively. Interestingly, in high crowded environments (more than 200 people), the impact of the proposed approach is impressive, proving to be the most effective solution; notably, even in less crowded environments (less than 200 people), the curriculum learning approach alone proves sufficient to manage the complexity of the environment, still overcoming the baseline approach.
There are many magnetic field harmonics in the operation of permanent magnet motor powered by frequency converter. In this case, the conventional coefficient fitting model has been unable to accurately calculate the iron loss of the motor. In this paper, a hybrid Jiles-Atherton hysteresis model is proposed. The inverse vector J-A model is used to describe the major hysteresis loop, and the inverse hyperbolic sine function is used to describe the minor hysteresis loop. Then the hysteresis loop of the core material of the prototype was mapped, and the parameters of the hybrid J-A model were identified by genetic algorithm. The finite element post-processing algorithm is written, and the iron loss of the sample motor is calculated. Finally, a test platform is set up to test motor loss under different working conditions. The difference of iron loss calculated by the hybrid J-A model and coefficient fitting model is analyzed and compared with the test value.
FeNi alloys exhibit outstanding soft magnetic properties, with wide applications in energy conversion and information processing technologies. While Mn doping and annealing have been shown to enhance magnetic performance in certain Fe-based alloys, its effect on the microstructure and magnetic properties of FeNi soft magnetically alloys remain poorly understood. This study systematically investigates the influence of different Mn doping and subsequent annealing on the microstructure and magnetic properties of as-forged FeNi alloys. Results show that higher Mn content in as-forged samples reduces permeability and increases coercivity, attributable to elevated stacking fault energy that suppresses recrystallization processes and maintains higher dislocation density, thereby impeding domain wall motion. Subsequent annealing significantly improves permeability while reducing coercivity with increasing annealing temperatures, due to dislocation annihilation, grain growth, and microstructural homogenization that facilitate domain wall movement and magnetic domain rotation. Notably, the higher Mn-doped alloys under identical annealing conditions exhibit superior magnetic performance, as the increased stacking fault energy inhibits annealing twin formation and reduces grain boundary pinning, thereby promoting grain boundary migration and microstructural homogenization. This work elucidates the regulatory mechanisms of Mn doping and annealing on the microstructure and magnetic properties of FeNi alloys, providing valuable insights for developing high-performance FeNi-based soft magnetic materials.
