Selected work

Representative Papers

SGAD: Semantic and Geometric-aware Descriptor for Local Feature Matching

Aug 04, 2025

Local feature matching suffers from low regional matching efficiency and heavy reliance on computationally expensive graph optimization. To address this, we propose the Semantic and Geometry-aware Descriptor Network (SGDN), an end-to-end trainable framework for direct region matching. SGDN introduces learnable regional descriptors, jointly supervised by classification and ranking losses, and incorporates a hierarchical containment-based redundancy filtering mechanism to enhance robustness. Geometric constraints are modeled via hierarchical containment graphs—bypassing explicit graph matching computation. Integrated with point-based matchers (e.g., LoFTR, ROMA), SGDN achieves both high accuracy and significant efficiency gains: 60× faster than MESA; outdoor pose estimation accuracy of 65.98 (surpassing DKM); and indoor AUC@5° improved by 7.39%, achieving state-of-the-art performance.

0 citationsRead paper

Traffic Signal Phase and Timing Estimation with Large-Scale Floating Car Data

Jul 13, 2025

Accurate estimation of Signal Phase and Timing (SPaT) remains challenging in the absence of inter-departmental coordination within transportation agencies. Method: This paper proposes the first scalable, highly robust, fully automated SPaT estimation framework, leveraging large-scale floating-car data. It integrates trajectory clustering, speed-change pattern recognition, adaptive time-series segmentation and denoising, and spatiotemporal contextual modeling to enable multi-period pattern identification and dynamic cycle detection—eliminating reliance on fixed signal cycles or simplified intersection topologies. Contribution/Results: The system processes over 15 million trajectories daily across more than two million traffic signals nationwide. It achieves SPaT estimation errors under 5 seconds for over 75% of signals and has been deployed in a production navigation platform, demonstrating industrial-grade scalability, robustness, and generalizability.

0 citationsRead paper

Multimodal HD Mapping for Intersections by Intelligent Roadside Units

Jul 11, 2025

To address the challenge of constructing high-definition semantic maps in complex intersections—where onboard solutions suffer from occlusions and limited field-of-view—this paper proposes a roadside infrastructure-based multimodal mapping method leveraging Intelligent Roadside Units (IRUs). We design a two-stage camera-LiDAR fusion framework that jointly optimizes modality-specific feature extraction and cross-modal semantic alignment, enabling high-fidelity geometric-textural joint modeling. We introduce RS-seq, the first publicly available sequential dataset specifically designed for roadside HD map generation, thereby establishing the first systematic benchmark for this emerging domain. Extensive evaluation on RS-seq demonstrates that our method achieves a semantic segmentation mIoU 4% higher than image-only baselines and 18% higher than point-cloud-only baselines, significantly outperforming existing approaches. This work establishes a new paradigm for high-precision, perception-driven semantic mapping from roadside infrastructure.

0 citationsRead paper

End-to-End Generation of City-Scale Vectorized Maps by Crowdsourced Vehicles

Jul 11, 2025

High-precision city-scale vector map construction faces dual bottlenecks: prohibitive cost of traditional LiDAR-based approaches and poor robustness of single-vehicle perception methods. Method: This paper proposes an end-to-end vectorized map generation framework leveraging crowdsourced, multi-vehicle, multi-temporal perception data. Its core innovation is the Trip-Aware Transformer architecture, which enables unified modeling and fusion of cross-vehicle and cross-temporal perception outputs via hierarchical spatiotemporal matching and multi-objective joint optimization. Contribution/Results: Evaluated on a large-scale real-world multi-city dataset, the method significantly outperforms single-vehicle baselines in vector map accuracy, with substantial improvements in structural completeness and geometric fidelity. It reduces manual annotation effort by 90%, achieving both high efficiency and strong generalization. The framework establishes a novel paradigm for low-cost, scalable, high-precision urban mapping.

0 citationsRead paper

Semantic SLAM with Rolling-Shutter Cameras and Low-Precision INS in Outdoor Environments

Apr 01, 2025

To address the low accuracy and severe drift in outdoor autonomous driving localization and mapping when fusing consumer-grade rolling-shutter cameras with low-cost inertial navigation systems (INS), this paper proposes the first real-time method integrating road-level semantic features—specifically lane markings and traffic signs—into a tightly coupled graph-optimization SLAM framework, jointly compensating for rolling-shutter distortion and long-term INS drift. Our approach unifies a semantic detection network, a rolling-shutter motion compensation model, multi-sensor (camera/IMU/wheel odometry) tight coupling in nonlinear optimization, and structural road priors as geometric constraints. Experiments demonstrate improvements of 5.35% in semantic detection recall and 2.79% in precision; relative pose error is bounded within 10 cm, and absolute positioning error remains approximately 1 m—even under challenging urban conditions—while maintaining robustness. The method significantly enhances production-grade localization performance on cost-constrained hardware.

0 citationsRead paper

Recent publications

Latest Papers

SGAD: Semantic and Geometric-aware Descriptor for Local Feature Matching

Aug 04, 2025

Local feature matching suffers from low regional matching efficiency and heavy reliance on computationally expensive graph optimization. To address this, we propose the Semantic and Geometry-aware Descriptor Network (SGDN), an end-to-end trainable framework for direct region matching. SGDN introduces learnable regional descriptors, jointly supervised by classification and ranking losses, and incorporates a hierarchical containment-based redundancy filtering mechanism to enhance robustness. Geometric constraints are modeled via hierarchical containment graphs—bypassing explicit graph matching computation. Integrated with point-based matchers (e.g., LoFTR, ROMA), SGDN achieves both high accuracy and significant efficiency gains: 60× faster than MESA; outdoor pose estimation accuracy of 65.98 (surpassing DKM); and indoor AUC@5° improved by 7.39%, achieving state-of-the-art performance.

0 citationsRead paper

Traffic Signal Phase and Timing Estimation with Large-Scale Floating Car Data

Jul 13, 2025

Accurate estimation of Signal Phase and Timing (SPaT) remains challenging in the absence of inter-departmental coordination within transportation agencies. Method: This paper proposes the first scalable, highly robust, fully automated SPaT estimation framework, leveraging large-scale floating-car data. It integrates trajectory clustering, speed-change pattern recognition, adaptive time-series segmentation and denoising, and spatiotemporal contextual modeling to enable multi-period pattern identification and dynamic cycle detection—eliminating reliance on fixed signal cycles or simplified intersection topologies. Contribution/Results: The system processes over 15 million trajectories daily across more than two million traffic signals nationwide. It achieves SPaT estimation errors under 5 seconds for over 75% of signals and has been deployed in a production navigation platform, demonstrating industrial-grade scalability, robustness, and generalizability.

0 citationsRead paper

Multimodal HD Mapping for Intersections by Intelligent Roadside Units

Jul 11, 2025

To address the challenge of constructing high-definition semantic maps in complex intersections—where onboard solutions suffer from occlusions and limited field-of-view—this paper proposes a roadside infrastructure-based multimodal mapping method leveraging Intelligent Roadside Units (IRUs). We design a two-stage camera-LiDAR fusion framework that jointly optimizes modality-specific feature extraction and cross-modal semantic alignment, enabling high-fidelity geometric-textural joint modeling. We introduce RS-seq, the first publicly available sequential dataset specifically designed for roadside HD map generation, thereby establishing the first systematic benchmark for this emerging domain. Extensive evaluation on RS-seq demonstrates that our method achieves a semantic segmentation mIoU 4% higher than image-only baselines and 18% higher than point-cloud-only baselines, significantly outperforming existing approaches. This work establishes a new paradigm for high-precision, perception-driven semantic mapping from roadside infrastructure.

0 citationsRead paper

End-to-End Generation of City-Scale Vectorized Maps by Crowdsourced Vehicles

Jul 11, 2025

High-precision city-scale vector map construction faces dual bottlenecks: prohibitive cost of traditional LiDAR-based approaches and poor robustness of single-vehicle perception methods. Method: This paper proposes an end-to-end vectorized map generation framework leveraging crowdsourced, multi-vehicle, multi-temporal perception data. Its core innovation is the Trip-Aware Transformer architecture, which enables unified modeling and fusion of cross-vehicle and cross-temporal perception outputs via hierarchical spatiotemporal matching and multi-objective joint optimization. Contribution/Results: Evaluated on a large-scale real-world multi-city dataset, the method significantly outperforms single-vehicle baselines in vector map accuracy, with substantial improvements in structural completeness and geometric fidelity. It reduces manual annotation effort by 90%, achieving both high efficiency and strong generalization. The framework establishes a novel paradigm for low-cost, scalable, high-precision urban mapping.

0 citationsRead paper

Semantic SLAM with Rolling-Shutter Cameras and Low-Precision INS in Outdoor Environments

Apr 01, 2025

To address the low accuracy and severe drift in outdoor autonomous driving localization and mapping when fusing consumer-grade rolling-shutter cameras with low-cost inertial navigation systems (INS), this paper proposes the first real-time method integrating road-level semantic features—specifically lane markings and traffic signs—into a tightly coupled graph-optimization SLAM framework, jointly compensating for rolling-shutter distortion and long-term INS drift. Our approach unifies a semantic detection network, a rolling-shutter motion compensation model, multi-sensor (camera/IMU/wheel odometry) tight coupling in nonlinear optimization, and structural road priors as geometric constraints. Experiments demonstrate improvements of 5.35% in semantic detection recall and 2.79% in precision; relative pose error is bounded within 10 cm, and absolute positioning error remains approximately 1 m—even under challenging urban conditions—while maintaining robustness. The method significantly enhances production-grade localization performance on cost-constrained hardware.

0 citationsRead paper