LiDAR has been widely used in autonomous driving systems to provide high-precision 3D geometric information about the vehicle’s surroundings for perception, localization, and path planning. LiDAR-based point cloud semantic segmentation is an important task with a critical real-time requirement. However, most of the existing convolutional neural network (CNN) models for 3D point cloud semantic segmentation are very complex and can hardly be processed at real-time on an embedded platform. In this study, a lightweight CNN structure was proposed for projection-based LiDAR point cloud semantic segmentation with only 1.9 M parameters that gave an 87% reduction comparing to the state-of-the-art networks. When evaluated on a GPU, the processing time was 38.5 ms per frame, and it achieved a 47.9% mIoU score on Semantic-KITTI dataset. In addition, the proposed CNN is targeted on an FPGA using an NVDLA architecture, which results in a 2.74x speedup over the GPU implementation with a 46 times improvement in terms of power efficiency.

The key to autonomous navigation in unmanned systems is the ability to recognize static and moving objects in the environment and to support the task of predicting the future state of the environment, avoiding collisions, and planning. However, because the existing 3D LiDAR point-cloud moving object segmentation (MOS) convolutional neural network (CNN) models are very complex and have large computation burden, it is difficult to perform real-time processing on embedded platforms. In this paper, we propose a lightweight MOS network structure based on LiDAR point-cloud sequence range images with only 2.3 M parameters, which is 66% less than the state-of-the-art network. When running on RTX 3090 GPU, the processing time is 35.82 ms per frame and it achieves an intersection-over-union(IoU) score of 51.3% on the SemanticKITTI dataset. In addition, the proposed CNN successfully runs the FPGA platform using an NVDLA-like hardware architecture, and the system achieves efficient and accurate moving-object segmentation of LiDAR point clouds at a speed of 32 fps, meeting the real-time requirements of autonomous vehicles.

The issue of obstacle avoidance and safety for visually impaired individuals has been a major topic of research. However, complex street environments still pose significant challenges for blind obstacle detection systems. Existing solutions often fail to provide real-time, accurate obstacle avoidance decisions. In this study, we propose a blind obstacle detection system based on the PC-CS-YOLO model. The system improves the backbone network by adopting the partial convolutional feed-forward network (PCFN) to reduce computational redundancy. Additionally, to enhance the network’s robustness in multi-scale feature fusion, we introduce the Cross-Scale Attention Fusion (CSAF) mechanism, which integrates features from different sensory domains to achieve superior performance. Compared to state-of-the-art networks, our system shows improvements of 2.0%, 3.9%, and 1.5% in precision, recall, and mAP50, respectively. When evaluated on a GPU, the inference speed is 20.6 ms, which is 15.3 ms faster than YOLO11, meeting the real-time requirements for blind obstacle avoidance systems.

Depression is a prevalent mental disorder that imposes a significant public health burden worldwide. Although multimodal detection methods have shown potential, existing techniques still face two critical bottlenecks: (i) insufficient integration of global patterns and local fluctuations in long-sequence modeling and (ii) static fusion strategies that fail to dynamically adapt to the complementarity and redundancy among modalities. To address these challenges, this paper proposes a dynamic multimodal depression detection framework, DynMultiDep, which combines multi-scale temporal modeling with an adaptive fusion mechanism. The core innovations of DynMultiDep lie in its Multi-scale Temporal Experts Module (MTEM) and Dynamic Multimodal Fusion module (DynMM). On one hand, MTEM employs Mamba experts to extract long-term trend features and utilizes local-window Transformers to capture short-term dynamic fluctuations, achieving adaptive fusion through a long-short routing mechanism. On the other hand, DynMM introduces modality-level and fusion-level dynamic decision-making, selecting critical modality paths and optimizing cross-modal interaction strategies based on input characteristics. The experimental results demonstrate that DynMultiDep outperforms existing state-of-the-art methods in detection performance on two widely used large-scale depression datasets.

Mitotic figures in tumor tissues are an important criterion for diagnosing malignant lesions, and physicians often search for the presence of mitosis in whole slide imaging (WSI). However, prolonged visual inspection by doctors may increase the likelihood of human error. With the advancement of deep learning, AI-based automatic cytopathological diagnosis has been increasingly applied in clinical settings. Nevertheless, existing diagnostic models often suffer from high computational costs and suboptimal detection accuracy. More importantly, when assessing cellular abnormalities, doctors frequently compare target cells with their surrounding cells—an aspect that current models fail to capture due to their lack of intercellular information modeling, leading to the loss of critical medical insights. To address these limitations, we conducted an in-depth analysis of existing models and propose an Inter–Intra Hypergraph Neural Network (II-HGNN). Our model introduces a block-based feature extraction mechanism to efficiently capture deep representations. Additionally, we leverage hypergraph convolutional networks to process both intracellular and intercellular information, leading to more precise diagnostic outcomes. We evaluate our model on publicly available datasets under varying imaging conditions, and experimental results demonstrate that our approach consistently outperforms baseline models in terms of accuracy.

Occlusion is a critical challenge in person re-identification (ReID), as partial visibility severely degrades feature discriminability and matching reliability. To address this issue, we propose a novel framework termed Visibility-Guided and Occlusion-Simulated Learning (VGOSL) for robust person ReID. The framework consists of two key components: a part-aware visibility modeling (PVM) module and an occlusion box simulation (OBS) module. The PVM module explicitly estimates part-level visibility reliability and adaptively reweights local features to guide global representation learning, enabling the model to emphasize informative regions while suppressing occluded ones. Meanwhile, the OBS module introduces structured occlusion box simulation during training to enhance robustness against realistic obstruction patterns through multi-branch supervision. Extensive experiments on Occluded-DukeMTMC, DukeMTMC-reID, Market-1501, Partial-ReID, and MSMT17 demonstrate that the proposed framework achieves competitive performance under both occluded and holistic settings. The source code has been publicly released on GitHub.