LLLinhui Li

Linhui Li

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Linhui Li has a total of 71 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Trajectory planning for autonomous driving on urban curvy roads: A search-resampling-enhanced iterative optimization framework

    by Weiwei Ren, Jing Lian, Linhui Li

    Autonomous vehicles face significant challenges in balancing scene applicability, trajectory quality, and computational efficiency when generating rapid, accurate, and optimal trajectories on urban curvy roads. The prevailing approach is to frame this task as an optimal control problem (OCP). However, the nominal OCP still faces problems due to its high-dimensional complexity. Although replacing large-scale collision avoidance constraints with driving corridor constraints reduces computational complexity, it inevitably sacrifices part of the free space and affects trajectory optimality. To address this issue, A Search-Resampling-Enhanced Iterative Optimization (SREIO) framework is proposed in this paper. The framework consists of three main stages: (1) Trajectory Search and Decision-Making, (2) Resampling Process, and (3) Enhanced Iterative Optimization. In the first stage, the dynamic programming (DP) algorithm is employed to find a coarse trajectory on three-dimensional space of S-L-T, defined by the longitudinal direction along the lane centerline (S), the lateral direction perpendicular to the lane centerline (L), and time (T). In the second stage, adjacent decision points are connected using a series of quintic polynomials and resampled along the resulting curve. After transforming the resampled trajectory from the Frenet frame to the Cartesian frame, it serves as the initial guess for the enhanced iterative optimization stage. In the third stage, a novel safe driving corridor (SDC) constraint is designed, which ensures collision avoidance by keeping the ego vehicle confined within the SDC. In addition, an intermediate OCP is constructed in each iteration process, where the current SDC is built based on the SDC from the previous iteration. Both simulation and real-vehicle experiments validate that the SREIO method exhibits significant advantages in scene applicability, trajectory quality, and computational efficiency. Notably, higher trajectory quality means safer and smoother planned trajectories.

  • DSRCF: Radar-camera fusion with dynamic filtering and coupled matching for object detection across driving scenarios

    by Nana Lyu, Jian Zhao, Pengbo Liu, Linhui Li, Jingbin Xu, Mingyan Wang, Yifan Zhang, Yunwei Zheng

    Multi-modal fusion plays a crucial role in enhancing the accuracy and robustness of autonomous driving perception systems. Due to the robust performance of millimeter-wave radar in adverse weather, the fusion of economical radar and camera sensors for object detection is an active research topic. However, the sparsity, multi-noise, and multipath interference of radar data pose significant challenges in effectively extracting and utilizing radar information, limiting the performance of the camera and radar fusion system. To address these issues, a novel fusion framework of radar and camera for efficient object detection across scenes is proposed, which achieves high detection accuracy while maintaining a fast inference speed. Specifically, a multi-dimensional dynamic filtering radar preprocessing method, which performs dynamic multi-target matching across consecutive frames, is developed to address discontinuity and low accuracy in radar feature extraction. During this period, the information expansion of the extracted radar features improves the contribution value of radar information within the fusion architecture. Besides, a coupled matching fusion strategy is implemented, enhancing the robustness of the proposed framework in cross-scenario detection tasks. Comprehensive experiments and ablation studies conducted on the nuScenes dataset validate the effectiveness of the proposed method, yielding a 4.8% improvement in cross-scenario mAP and up to a 7.5% improvement in single-scenario mAP.

  • Multi-sensor tightly coupled localization and semantic mapping based on adaptive voxel

    by Shi Chen, Jing Lian, Qiuyu Zang, Duo Sui, Jian Zhao, Linhui Li


  • Hierarchical Recursive Interaction and Multi-Stage Goal-Guided Mechanism for Multimodal Trajectory Prediction

    by Xuecheng Wang, Linhui Li, Jing Lian, Zhenfeng Wang, Juan Li, Jian Zhao, Qiong Wu, Jun Hu

    In highly dynamic and complex autonomous driving environments, accurately predicting agents’ future multimodal trajectories still faces challenges such as modeling diverse social interactions, capturing dynamic intents, and ensuring prediction consistency. To address these issues, this paper proposes a novel trajectory prediction model that integrates a Hierarchical Recursive Interaction Network (HRINet) and a multi-stage goal-guided mechanism (GoalNet), aiming to improve prediction accuracy, stability, and plausibility. Specifically, we design a HRINet with local and global attention mechanisms to recursively model various social interactions, while progressively integrating map semantic information to enhance the model’s understanding of traffic scenes. Meanwhile, inspired by the divide-and-conquer approach, the proposed GoalNet first estimates fine-grained multi-stage goal lane segments along the path. These goals are then used to continuously guide and constrain the trajectory generation process, effectively reducing error accumulation and improving stability. In addition, we construct a dynamic goal candidate area that combines domain knowledge and traffic rules to filter out unreasonable goals, thereby enhancing the plausibility and consistency of the predictions. Experimental results on nuScenes, INTERACTION, and Waymo Open Motion Dataset (WOMD) show that our model achieves state-of-the-art performance in multiple key metrics, maintains a trade-off between prediction accuracy, model complexity, and inference latency, and shows high stability and consistency in predictions.


  • RCQFusion: Radar and Camera Sensor Fusion with Joint Quantization for Robust Object Detection in Autonomous Driving

    by Nana Lyu, Jian Zhao, Pengbo Liu, Linhui Li, Chunlin Li

    Object detection constitutes a crucial component for environment perception in autonomous driving systems, and the multi-modal fusion of millimeter-wave radar and camera is an effective solution to improve the accuracy and robustness of object detection. However, inherent sparsity and multi-noise of automotive radar data pose challenges to feature extraction, thereby resulting in low radar utilization. This limitation is further evidenced by existing fusion methods, where the decision weight assignment between the camera and the radar is unbalanced, and the decision contribution of radar often remains suboptimal in fusion processes. To address these issues, a novel radar-camera fusion framework with joint quantization for object detection is proposed. Specifically, considering the non-uniform density distribution and spatial sparsity of radar point clouds, an improved multi-dimensional density clustering algorithm is first developed to achieve effective noise filtering. Subsequently, a radar confidence factor is introduced based on the target point cloud's scalability to expand radar features. Furthermore, a dual-channel joint quantization matching mechanism is devised to enhance the decision contribution of radar during the fusion process. Comprehensive experiment and ablation studies conducted on the nuScenes dataset demonstrate that the proposed method exhibits superior inference speed, and the mAP outperforms comparison methods.


  • Robust Trajectory Prediction Based on Efficient Fusion of Heterogeneous Information

    by Xuecheng Wang, Dongfang Yang, Linhui Li, Chunlin Li, Jing Lian, Jian Zhao, Zhongxu Zhang, Yicheng Bian

    Trajectory prediction plays a crucial role in enhancing risk warning and scene understanding capabilities in autonomous driving systems. Despite significant advancements in recent years, challenges remain in balancing real-time inference and prediction accuracy, suppressing data noise, and reducing reliance on high-definition maps. To address these issues, this study proposes a robust trajectory prediction model that efficiently fuses heterogeneous information. First, we design a novel compact vector representation method to explicitly encode the spatial relationships between map elements and agents, thereby improving the model's ability to comprehend traffic scenarios. Subsequently, we propose an efficient and reusable information processing and fusion mechanism, the Stacked Interaction-aware Module (SIM), to extract heterogeneous information and aggregate various interaction features. SIM not only enhances information processing efficiency but also improves robustness against noisy data due to its modular and reusable design. Finally, encoding the historical states of traffic agents with multiple strategies, the model can still achieve effective trajectory prediction even without relying on map information. Experimental results demonstrate that the proposed model outperforms existing state-of-the-art methods on multiple public and internal datasets. Furthermore, it has been successfully deployed in real-world autonomous driving systems, achieving efficient operation at 90 frames per second on an embedded GPU.

  • Open Access

    Union of Experts: Adapting Hierarchical Routing to Equivalently Decomposed Transformer

    by Yujiao Yang Yang, Jing Lian, Linhui Li

    Mixture-of-Experts (MoE) enhances model performance while maintaining computational efficiency, making it well-suited for large-scale applications. In existing MoE paradigms, each expert works as an individual, leading to inadequate collaboration. Moreover, the MoE framework has not been effectively extended to attention blocks, which limits further efficiency improvements. To tackle these issues, we propose Union-of-Experts (UoE), which decomposes the transformer model into an equivalent group of experts and applies selective routing to input data and experts. Our approach advances MoE design with four key innovations: (1) We conduct equivalent expert decomposition on both MLP blocks and attention blocks based on matrix partitioning in tensor parallelism. (2) We develop two routing paradigms: patch-wise data selection and expert selection, to apply routing at different levels. (3) We design the architecture of the UoE model, including Selective Multi-Head Attention (SMHA) and Union-of-MLP-Experts (UoME). (4) We develop the parallel implementation of UoE’s routing and computation operations and optimize the efficiency based on hardware processing analysis. The experiments demonstrate that our UoE model surpasses Full Attention, state-of-the-art MoEs, and efficient transformers (including the recently proposed DeepSeek-V3 architecture) in several tasks across image and natural language domains. In language modeling tasks, UoE achieves an average reduction of 2.38 in perplexity compared to the best-performing MoE method with only 76% of its FLOPs. In the Long Range Arena benchmark, it demonstrates an average score at least 0.68% higher than all comparison models, including Full Attention, MoEs, and transformer variants, with only 50% of the FLOPs of the best MoE method. In image classification, our model yields an average accuracy improvement of 1.75% over the best model while maintaining comparable FLOPs. The source codes are available at https://github.com/YujiaoYang-work/UoE.


  • Dual-STGAT: Dual Spatio-Temporal Graph Attention Networks With Feature Fusion for Pedestrian Crossing Intention Prediction

    by Jing Lian, Yiyang Luo, Xuecheng Wang, Linhui Li, Ge Guo, Weiwei Ren, Tao Zhang

    Pedestrian intent prediction is critical for autonomous driving, as accurately predicting crossing intentions helps prevent collisions and ensures the safety of both pedestrians and passengers. Recent research has focused on vision-based deep neural networks for this task, but challenges remain. First, current methods suffer from low efficiency in multi-feature fusion and unreliable predictions under challenging conditions. Additionally, real-time performance is essential in practical applications, so the efficiency of the algorithm is crucial. To address these issues, we propose a novel architecture, Dual-STGAT, which uses a dual-level spatio-temporal graph network to extract pedestrian pose and scene interaction features, reducing information loss and improving feature fusion efficiency. The model captures key features of pedestrian behavior and the surrounding environment through two modules: the Pedestrian Module and the Scene Module. The Pedestrian Module extracts pedestrian motion features using a spatio-temporal graph attention network, while the Scene Module models interactions between pedestrians and surrounding objects by integrating visual, semantic, and motion information through a graph network. Extensive experiments conducted on the PIE and JAAD datasets show that Dual-STGAT achieves over 90% accuracy in pedestrian crossing intention prediction, with inference latency close to 5ms, making it well-suited for large-scale production autonomous driving systems that demand both performance and computational efficiency.