Failure to wear seatbelts remains one of the primary factors contributing to severe injuries and fatalities in traffic accidents, making robust in-vehicle seatbelt detection important for intelligent cockpit safety monitoring. Existing vision-based methods are predominantly adapted from general object detection frameworks and may struggle when seatbelts are slender, low-contrast, or visually similar to clothing and seat backgrounds. To improve seatbelt wearing-status recognition under such complex in-vehicle conditions, this paper proposes KPGBeltNet, a seatbelt detection algorithm that integrates human keypoint priors with geometric feature modeling. The proposed method adopts a two-stage detection strategy: firstly, YOLOv11-pose is employed to extract human keypoints, including shoulders and hips, enabling more precise and pose-robust region of interest localization than conventional whole-body bounding box detection; secondly, a block diagonal sampling strategy is introduced to extract a sequence of overlapping local image patches along the shoulder-to-hip direction, thereby capturing the geometric distribution of seatbelts while reducing irrelevant background interference. A local–global attention mechanism is then designed to model the interaction between global region features and local patch features, injecting contextual information into local representations to enhance discriminative capability. Finally, a Bidirectional GRU (Bi-GRU) network is employed to model dependencies among the sampled patch sequence and generate comprehensive feature representations for classification. Experimental results on a self-constructed in-vehicle monitoring dataset demonstrate that KPGBeltNet achieves superior detection accuracy and robustness under complex conditions, including illumination variation and passenger pose changes, providing effective technical support for intelligent cockpit safety monitoring systems. Code and datasets for this paper are available at https://github.com/tjc609/KPGBeltNet .

Shadow segmentation is sensitive to illumination shifts and scene diversity. Test-time adaptation provides a source-free solution, yet most methods assume large batches or stationary distributions, limiting online use. We present an online test-time adaptation framework for shadow segmentation, where data arrive sequentially in small batches without revisiting. Our method applies dual-stage filtering: image-level domain scoring with a frozen supervisor, followed by pixel-level soft masking guided by confidence, consistency, and structural cues. A brightness-guided module exploits illumination priors to refine shadow regions, while curriculum-based thresholding and exponential moving average teacher–student updates stabilize adaptation and reduce error accumulation. Experiments across six cross-domain settings on SBU, CUHK-Shadow, and ISTD demonstrate consistent gains over baselines in accuracy and boundary preservation, achieving robust generalization under diverse shifts.

With the advancement of vision-based autonomous driving technology, pedestrian detection have become an important component for improving traffic safety and driving system robustness. Nevertheless, in complex traffic scenarios, conventional pose estimation approaches frequently fail to accurately reconstruct occluded keypoints, primarily due to obstructions caused by vehicles, vegetation, or architectural elements. To address this issue, we propose a novel real-time occluded pedestrian pose completion framework termed Separation and Dimensionality Reduction-based Generative Adversarial Imputation Nets (SDR-GAIN). Unlike previous approaches that train visual models to distinguish occlusion patterns, SDR-GAIN aims to learn human pose directly from the numerical distribution of keypoint coordinates and interpolate missing positions. It employs a self-supervised adversarial learning paradigm to train lightweight generators with residual structures for the imputation of missing pose keypoints. Additionally, it integrates multiple pose standardization techniques to alleviate the difficulty of the learning process. Experiments conducted on the COCO and JAAD datasets demonstrate that SDR-GAIN surpasses conventional machine learning and Transformer-based missing data interpolation algorithms in accurately recovering occluded pedestrian keypoints, while simultaneously achieving microsecond-level real-time inference.

In recent years, intelligent vehicles operating in urban environments have demonstrated the capability to autonomously execute various tasks, such as object detection, lane detection, segmentation, etc. This advancement is facilitated by the extensive datasets accumulated by researchers, alongside advancements in intelligent algorithms, as well as significant breakthroughs in software and hardware. However, within the autonomous driving community, there is a scarcity of data regarding scenarios encountered in mining environments. This scarcity presents challenges and bottlenecks for the advancement of comprehensive autonomous driving systems and autonomousoperations. Although we previously released our dataset, AutoMine, which includes over 18 hours of driving data in open-pit mines, its scope is limited to two specific tasks. This scope limitation impedes the training and validation of the majority of algorithms for different tasks in this particular scenario. To broaden the scope of autonomous driving visual tasks in mining environments, we have curated a diverse collection encompassing multiple tasks, including detection, segmentation, tracking, etc. Additionally, we have established benchmarks and set up baselines for the aforementioned multiple tasks. By comparing the performance differences of visual algorithms between mining areas and other scenarios, we demonstrate the distinctive characteristics of mining regions in an intuitive manner. We have developed a suite of tools for converting annotated data into the standardized format used in existing driving datasets. Our aspiration is to establish data and benchmark foundations, supporting research endeavors in intelligent transportation within mining environments and autonomous driving in comprehensive scenarios. Our project website can be seen in AutoMine, and the dataset can be downloaded via AutoMine-Benchmark.

This paper presents a unified two-stage framework for physics-based musculoskeletal motion control and generation. To tackle the challenges posed by high dimensionality and redundancy, we first employ an autoencoder to learn a low-dimensional muscle synergy space from activation data. Policies trained in this space make the character faithfully reproduce motions while generating physiologically plausible muscle activations. We then leverage these expert trajectories to train a Conditional VAE, encoding skills into a continuous latent space for task-agnostic motion synthesis and downstream control. Experiments show our method achieves high motion imitation accuracy and generation diversity, ensures control stability, and maintains physiological realism, offering an effective solution for generalizing control of complex musculoskeletal characters.

Local feature matching plays a critical role in robotic SLAM and visual localization. However, in weakly textured indoor industrial environments, lightweight appearance-based methods often struggle to learn discriminative and stable local features. To address this challenge, this paper proposes GAEFeat, short for Geometry-Aware Efficient Feature, a lightweight vision-geometric feature learning network. To address the scarcity of specialized training data, we integrated robotic arm pose priors with depth information to automatically generate cross-view supervision signals and surface-normal labels. Based on this strategy, we constructed two complementary datasets, including a simulated dataset and a real-world dataset, to support feature learning and evaluation in weakly textured indoor industrial environments. For feature extraction, we design a dual enhancement mechanism consisting of a geometric auxiliary branch and a geometry-aware enhancement (GAE) module. The former guides the network to perceive local surface structures through surface normal supervision, while the latter utilizes a gating mechanism to achieve deep fusion between geometric priors and 2D texture descriptors. Experimental results demonstrate that GAEFeat achieves strong robustness and high inference efficiency in relative pose estimation, homography estimation, and visual localization tasks, with particularly notable advantages in near-field, weakly textured industrial scenes. The framework achieves an inference latency of only 3.9 ms on the NVIDIA Jetson AGX Orin edge platform, demonstrating its real-time capability and practical potential for deployment in edge computing environments.

Depth priors are critical for maintaining accurate geometry in novel view synthesis, especially when input views are sparse and scenes are complex and room-scale. However, obtaining reliable depth under such conditions remains a significant challenge. In this work, we present a novel method that self-supervisedly adapts a pretrained video depth estimation network to each specific scene. This scene-specific adaptation enables high-fidelity depth estimation even under extreme view sparsity. Leveraging these accurate depth priors, we propose an efficient two-stage NeRF training pipeline with a density self-correction mechanism that mitigates the effects of depth errors without relying on explicit uncertainty modeling. Our approach not only enhances synthesis quality in sparse-view settings but also significantly accelerates neural radiance field (NeRF) training, outperforming existing methods in both fidelity and efficiency.

Driver drowsiness poses a critical threat to road safety and remains a persistent challenge even in advanced autonomous driving systems. However, existing vision-based detection methods extract facial features inefficiently and are susceptible to cumulative errors from external face alignment models. To address these limitations, we develop a novel video-based framework that combines trainable facial feature extraction with efficient spatial-temporal analysis. Our feature extraction module employs stacked landmark heatmaps to extract structural facial features efficiently while can be trained for drowsiness-specific features. Furthermore, we introduce a diagonal embedding strategy to enhance the geometric relations between landmarks of a single stacked heatmap. To adequately analyze the spatial-temporal features, a two-stream structure combining 2D and 3D convolutions with multi-head self-attention is utilized in this article. Experimental results validate that our method outperforms existing methods and achieves higher detection accuracy.

Semantic segmentation is of great importance in the field of autonomous driving, as it provides semantic information for a scene that intelligent vehicles need to interact with. Although a large number of different semantic segmentation networks have been proposed, achieving high performance for semantic segmentation in real-time using a lightweight network is challenging in practical conditions. In this article, we propose an efficient postprocessing network that can be applied to various real-time semantic segmentation networks to enhance their performance. Specifically, we introduce a transformer-based lightweight network to obtain information for refining the output of a given semantic segmentation network. Our network has very limited parameters and can work in real-time and a plug-and-play manner to enhance the performance of different semantic segmentation networks. This capability can significantly benefit real-time perception in autonomous driving applications. We demonstrate the effectiveness of our network through extensive experiments showing that it can improve the performance of various semantic segmentation networks.

3D object detection is essential for robust environmental perception in autonomous driving and robotics. While LiDAR-camera fusion methods offer high accuracy, their computational complexity hinders deployment on resource-constrained edge devices. To address this, we introduce QuantBEVFusion, a fully quantized 3D object detection framework that prioritizes both quantization and operator optimization. Our approach tackles the inherent asymmetry between LiDAR and camera data in pillar-based models by incorporating a novel pillar bird’s-eye view (BEV) encoder, significantly boosting performance. Furthermore, we introduce 1) an optimized LiDAR input processing method that filters noise and enables per-tensor quantization; 2) an improved sparse feature quantization process with log-histogram balancing, adaptive bin widths, and distillation loss for enhanced accuracy; and 3) a deployment-friendly 3D-to-2D transformation operator facilitating fixed-point implementation. Extensive experiments demonstrate that QuantBEVFusion achieves state-of-the-art quantization performance while maintaining accuracy suitable for real-time applications on edge devices.

Recent methods, such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have demonstrated remarkable capabilities in novel view synthesis. However, despite their success in producing high-quality images for viewpoints similar to those seen during training, they struggle when generating detailed images from viewpoints that significantly deviate from the training set, particularly in close-up views. The primary challenge stems from the lack of specific training data for close-up views, leading to the inability of current methods to render these views accurately. To address this issue, we introduce a novel pseudo-label-based learning strategy. This approach leverages pseudo-labels derived from existing training data to provide targeted supervision across a wide range of close-up viewpoints. Recognizing the absence of benchmarks for this specific challenge, we also present a new dataset designed to assess the effectiveness of both current and future methods in this area. Our extensive experiments demonstrate the efficacy of our approach.

Visual odometry (VO), including keypoint detection, correspondence establishment, and pose estimation, is a crucial technique for determining motion in machine vision, with significant applications in augmented reality (AR), autonomous driving, and visual simultaneous localization and mapping (SLAM). For feature-based VO, the repeatability of keypoints affects the pose estimation. The convolutional neural network (CNN)-based detectors extract high-level features from images, thereby exhibiting robustness to viewpoint and illumination changes. Compared with descriptor matching, optical flow tracking exhibits better real-time performance. However, mainstream CNN-based detectors rely on the “joint detection and descriptor” framework to realize matching, making them incompatible with optical flow tracking. To obtain keypoints suitable for optical flow tracking, we propose a self-supervised detector based on transfer learning named OFPoint, which jointly calculates pixel-level positions and confidences. We use the descriptor-based detector simple learned keypoints (SiLK) as the pre-trained model and fine-tune it to avoid training from scratch. To achieve multi-scale feature fusion in detection, we integrate the multi-scale attention mechanism. Furthermore, we introduce the maximum discriminative probability loss term, ensuring the grayscale consistency and local stability of keypoints. OFPoint achieves a balance between accuracy and real-time performance when establishing correspondences on HPatches. Additionally, we demonstrate its effectiveness in VO and its potential for graphics applications such as AR.

In domains such as autonomous driving and remote sensing, images often suffer from challenging lighting conditions, including low-light, backlighting and overexposure, which hinder the recognition of pedestrians, vehicles and traffic signs. While numerous methods have been proposed to address poor image exposure, they often struggle with images containing both low-light and overexposed regions. This paper presents an unsupervised learning-based exposure control method, providing a novel approach to improving image quality under diverse lighting conditions. Leveraging the inherent properties of Retinex theory, we introduce a novel yet simple formula that adjusts image exposure to produce visually pleasing results without requiring paired training data. Experiments on diverse image datasets validate the effectiveness of our approach in addressing various exposure challenges while preserving critical visual details. Our framework not only simplifies the exposure control process but also achieves state-of-the-art performance, highlighting its potential for real-world applications in computer vision and image processing.

In the field of computer graphics, crowd simulation is crucial for constructing realistic virtual environments. This paper introduces a novel approach to adaptive formation control and transformation in virtual crowds using deep reinforcement learning. Our method enables virtual agents to autonomously generate, maintain, and adaptively transform formations relying solely on local perception and interaction. The proposed framework comprises three key components: dynamic formation configuration based on alignment, a formation control algorithm embedded with formation constraints, and an adaptive formation transformation mechanism. Simulation results demonstrate that our method achieves formation errors around 5 https://github.com/SYDDX/AFCTVC-DRL .

Effective perception and accurate localization of lane lines are the key points for intelligent vehicles to plan local driving paths and realize lane keeping and departure warning. However, the elongated structure of lane lines makes the performance of detectors degrade significantly when visual cues are scarce. The continuity of lane lines also puts forward higher requirements for the ability of algorithms to model long-range dependencies. In this paper, we propose a novel anchor-based lane detection network (SP-Det) combining the unique structural characteristics and pixel distribution of lane lines. Specifically, we introduce a Semantic-Guided Feature Calibration Unit (SG-FCU) to semantically calibrate and refine features from different layers and to narrow the semantic gap during fusion. Additionally, we propose a Spatial-aware Context Aggregation Block (S-CAB) and a Lane-aware Information Enhancement Module (LIEM) to improve the prediction accuracy of horizontal offsets of line anchors through global feature encoding and row-wise information sharing. The results of quantitative and qualitative experiments show that SP-Det achieves state-of-the-art performance on CULane and Tusimple benchmark datasets.

The body condition scoring (BCS) of pigs comprehensively reflects the health status of pigs by evaluating their fat and muscle reserves. Traditional evaluation methods rely on manual measurement of pig backfat thickness, which is inefficient and can causes severe stress reactions in pigs. In recent years, the development of computer vision has provided new ideas for pig BCS evaluation. This paper proposes a pig body condition score evaluation method based on the MobileSAM. The method first detects pigs from a rear-view perspective, then uses the resulting bounding boxes as prompts for an improved MobileSAM to perform instance segmentation. We enhanced MobileSAM's segmentation accuracy by fine-tuning the model, adapting MobileSAM from general instance segmentation tasks to the specific task of pig instance segmentation. Finally, we calculated the aspect ratio (LWR) of the minimum bounding rectangle of the mask from the pig's rear-view perspective. A decision tree model is employed to perform a non-linear mapping of the LWR value, classifying it into one of five predefined body condition score levels to determine the final grade. Our method effectively overcomes the limitations of traditional manual measurements. It features low computational cost, excellent realtime performance, and supports deployment on edge computing terminals in environments such as pig farms. Experimental results show that our method has significant advantages over other image classification and object detection methods in terms of accuracy and interpretability.

The development of the systems capable of synthesizing natural and life-like motions for virtual characters has long been a central focus in computer animation. It needs to generate high-quality motions for characters and provide users with a convenient and flexible interface for guiding character motions. In this work, we propose a language-directed virtual human motion generation approach based on musculoskeletal models to achieve interactive and higher-fidelity virtual human motion, which lays the foundation for the development of language-directed controllers in physics-based character animation. First, we construct a simplified model of musculoskeletal dynamics for the virtual character. Subsequently, we propose a hierarchical control framework consisting of a trajectory tracking layer and a muscle control layer, obtaining the optimal control policy for imitating the reference motions through the training. We design a multi-policy aggregation controller based on large language models, which selects the motion policy with the highest similarity to user text commands from the action-caption data pool, facilitating natural language-based control of virtual character motions. Experimental results demonstrate that the proposed approach not only generates high-quality motions highly resembling reference motions but also enables users to effectively guide virtual characters to perform various motions via natural language instructions. We propose a language-directed virtual human motion generation approach based on musculoskeletal models to achieve interactive and higher-fidelity virtual human motion. It takes reference motion data, caption and text prompts as inputs, realizing the natural language motion controller through three components: constructing an action-caption data pool, learning the control policies for imitating the motion, and semantic matching selection. image

Formation maintenance and collision avoidance allow multiple agents to form and maintain a specific formation in the environment and reach the final goal safely without any collisions, which are important component of the multi-agent/crowd cooperation field. We propose a two-level formation control method to simulate crowd formation. At the low level, a comprehensive group reward function is designed to train the agents to learn to collaborate to reach the goal and avoid collisions with other agents and obstacles. The Multi-Agent POsthumous Credit Assignment algorithm and curriculum learning are combined to solve the credit assignment problem in the shared group reward and the slow convergence problem of the policy in the reward-sparse environment. At the high level, we design a formation switch method to determine the compression or splitting of the formation for multiple agents passing through the narrow exit. Then the optimal reciprocal collision avoidance algorithm is adopted to compute a collision-free velocity for the team of multiple agents, on the basis of which the desired formation positions are calculated to avoid collisions between the team and the obstacles in the environment. Lastly, Hungarian algorithm is utilized to assign the formation positions to the agents as their respective target position to accelerate the process of forming. Experimental results demonstrate that our proposed approach enable multiple agents to constitute the formation shape from the starting positions quickly and maintain the shape well when moving toward the final goal. Furthermore, our model has a certain generalization ability to unseen formation shapes in the scenarios without and with obstacles.

Recent advances in discriminative and generative pretraining have yielded geometry estimation models with strong generalization capabilities. While discriminative monocular geometry estimation methods rely on large-scale fine-tuning data to achieve zero-shot generalization, several generative-based paradigms show the potential of achieving impressive generalization performance on unseen scenes by leveraging pre-trained diffusion models and fine-tuning on even a small scale of synthetic training data. Frustratingly, these models are trained with different recipes on different datasets, making it hard to find out the critical factors that determine the evaluation performance. Besides, current geometry evaluation benchmarks have two main drawbacks that may prevent the development of the field, i.e., limited scene diversity and unfavorable label quality. To resolve the above issues, (1) we build fair and strong baselines in a unified codebase for evaluating and analyzing the geometry estimation models; (2) we evaluate monocular geometry estimators on more challenging benchmarks for geometry estimation task with diverse scenes and high-quality annotations. Our results reveal that pre-trained using large data, discriminative models such as DINOv2, can outperform generative counterparts with a small amount of high-quality synthetic data under the same training configuration, which suggests that fine-tuning data quality is a more important factor than the data scale and model architecture. Our observation also raises a question: if simply fine-tuning a general vision model such as DINOv2 using a small amount of synthetic depth data produces SOTA results, do we really need complex generative models for depth estimation? We believe this work can propel advancements in geometry estimation tasks as well as a wide range of downstream applications.

Existing neural network models directly add attention mechanisms to the network as a plug-and-play component to capture long-range dependencies and reconstruct feature maps. However, most methods do not fully tap the potential of attention in dealing with multi-scale problems. In this paper, an attention-guided YOLOv4 network (AG-YOLO) is proposed to address the multi-scale issue in object detection. We propose and apply multi-scale feature extraction to later stages of the backbone, which can not only enrich the feature hierarchy with low computational overhead, but also model the intra-scale and inter-scale correlation simultaneously to avoid missing key information. To reduce the redundant use of information flow, we propose a lightweight attention-guided feature pyramid network, which provides an efficient multi-level aggregation strategy based on multi-scale channel attention. In addition, a global context pathway is designed to reduce the dilution of high-level semantic information caused by information transmission. Compared with the baseline, AG-YOLO increased the mAP_0.5 by 1.67%, while the number of parameters and GFLOPs merely increased by 0.33M and 0.18, respectively. Meanwhile, the detection accuracy of small object categories has been improved.