Location-Aware Language Models via Secondary Embeddings
该研究提出了一种轻量级方法，通过向预训练的语言模型中注入地理空间信息来改善地点名称和空间实体的表示，无需重新训练或修改分词器。
0 citationsRead paper
Selected work
Representative Papers
Anchor-Based AI Approach for Pre-Crash Object Detection Utilizing Micro-Doppler Signatures in Automotive Radar
This study addresses the challenges of missed detections and inaccurate motion parameter estimation for near-field targets in automotive pre-crash scenarios by proposing an anchor-based deep learning model tailored for high-resolution automotive radar data. The method innovatively incorporates radar image dilation into the input feature channels to enhance local micro-Doppler signatures, effectively mitigating issues caused by sparse point clouds, signal fluctuations, and multipath interference. Experimental results demonstrate that the proposed model significantly outperforms conventional radar tracking approaches on a real-world pre-crash dataset, exhibiting superior generalization capability and higher detection reliability in dynamic environments. These improvements provide robust perceptual support for intelligent safety systems, such as pretensioner airbags, enabling more accurate and timely responses.
0 citationsRead paper
Hardware-triggered Time Synchronization of Roadside Multi-lidar, Multi-camera Measurement System for Accurate Data Alignment
This work addresses the challenge of time synchronization among heterogeneous sensors in roadside and vehicle-mounted multi-LiDAR–multi-camera systems by proposing an open-source, modular, and scalable hardware synchronization solution. Using the LiDAR synchronization pulse as a reference, the system employs programmable delay circuits to generate independent trigger signals for each camera, enabling flexible and precise spatiotemporal alignment. The architecture supports arbitrary combinations of sensor counts and has been validated on both a three-camera roadside platform and a seven-camera vehicular setup. Experimental results demonstrate significantly improved spatial consistency between point clouds and images, while the design ensures robustness, reproducibility, and ease of deployment.
0 citationsRead paper
Preserve the Hard, Regenerate the Rest: Uncertainty-Guided Synthetic Training Data Augmentation with Diffusion Models
This work addresses the performance limitations of semantic segmentation models on sparse and visually diverse rare regions—such as small objects in aerial imagery or autonomous driving scenes—where conventional synthetic augmentation often causes pixel-label misalignment and inefficient computation. The authors propose an uncertainty-guided contextual augmentation method that identifies uncertain regions via prediction entropy and selectively applies diffusion-based inpainting only to their complementary context. During fine-tuning, loss is computed exclusively on original pixels to preserve label consistency while maximizing information gain. Notably, this approach requires no external models or heuristic rules and represents the first integration of predictive uncertainty with diffusion-based restoration for targeted enhancement of challenging regions. Experiments on Cityscapes, UAVID, and BDD100K demonstrate significant mIoU improvements, with the largest gains observed in difficult classes such as buses, trains, and aerial-view cars.
0 citationsRead paper
RoAd-RL: A Unified Library and Benchmark for Robust Adversarial Reinforcement Learning
Deep reinforcement learning exhibits significantly degraded robustness under adversarial perturbations, and existing research suffers from fragmented implementations, inconsistent evaluation protocols, and poor reproducibility. This work proposes the first standardized, open-source benchmarking framework for adversarial reinforcement learning, which unifies abstract interfaces for policies, attacks, defenses, and robustness metrics, and seamlessly integrates with Stable-Baselines3 and Gymnasium. Supporting algorithms such as DQN, PPO, and SAC, the framework encompasses 192 attack–defense combinations. Systematic evaluations on LunarLander and Highway-v0 reveal substantial environment-dependent variations in agent robustness, demonstrate that certain defense methods can be detrimental, and consistently identify temporal smoothing as an effective strategy for enhancing robustness. The code is publicly released.
0 citationsRead paper
Recent publications
Latest Papers
Location-Aware Language Models via Secondary Embeddings
该研究提出了一种轻量级方法，通过向预训练的语言模型中注入地理空间信息来改善地点名称和空间实体的表示，无需重新训练或修改分词器。
0 citationsRead paper
Anchor-Based AI Approach for Pre-Crash Object Detection Utilizing Micro-Doppler Signatures in Automotive Radar
This study addresses the challenges of missed detections and inaccurate motion parameter estimation for near-field targets in automotive pre-crash scenarios by proposing an anchor-based deep learning model tailored for high-resolution automotive radar data. The method innovatively incorporates radar image dilation into the input feature channels to enhance local micro-Doppler signatures, effectively mitigating issues caused by sparse point clouds, signal fluctuations, and multipath interference. Experimental results demonstrate that the proposed model significantly outperforms conventional radar tracking approaches on a real-world pre-crash dataset, exhibiting superior generalization capability and higher detection reliability in dynamic environments. These improvements provide robust perceptual support for intelligent safety systems, such as pretensioner airbags, enabling more accurate and timely responses.
0 citationsRead paper
Hardware-triggered Time Synchronization of Roadside Multi-lidar, Multi-camera Measurement System for Accurate Data Alignment
This work addresses the challenge of time synchronization among heterogeneous sensors in roadside and vehicle-mounted multi-LiDAR–multi-camera systems by proposing an open-source, modular, and scalable hardware synchronization solution. Using the LiDAR synchronization pulse as a reference, the system employs programmable delay circuits to generate independent trigger signals for each camera, enabling flexible and precise spatiotemporal alignment. The architecture supports arbitrary combinations of sensor counts and has been validated on both a three-camera roadside platform and a seven-camera vehicular setup. Experimental results demonstrate significantly improved spatial consistency between point clouds and images, while the design ensures robustness, reproducibility, and ease of deployment.
0 citationsRead paper
Preserve the Hard, Regenerate the Rest: Uncertainty-Guided Synthetic Training Data Augmentation with Diffusion Models
This work addresses the performance limitations of semantic segmentation models on sparse and visually diverse rare regions—such as small objects in aerial imagery or autonomous driving scenes—where conventional synthetic augmentation often causes pixel-label misalignment and inefficient computation. The authors propose an uncertainty-guided contextual augmentation method that identifies uncertain regions via prediction entropy and selectively applies diffusion-based inpainting only to their complementary context. During fine-tuning, loss is computed exclusively on original pixels to preserve label consistency while maximizing information gain. Notably, this approach requires no external models or heuristic rules and represents the first integration of predictive uncertainty with diffusion-based restoration for targeted enhancement of challenging regions. Experiments on Cityscapes, UAVID, and BDD100K demonstrate significant mIoU improvements, with the largest gains observed in difficult classes such as buses, trains, and aerial-view cars.
0 citationsRead paper
RoAd-RL: A Unified Library and Benchmark for Robust Adversarial Reinforcement Learning
Deep reinforcement learning exhibits significantly degraded robustness under adversarial perturbations, and existing research suffers from fragmented implementations, inconsistent evaluation protocols, and poor reproducibility. This work proposes the first standardized, open-source benchmarking framework for adversarial reinforcement learning, which unifies abstract interfaces for policies, attacks, defenses, and robustness metrics, and seamlessly integrates with Stable-Baselines3 and Gymnasium. Supporting algorithms such as DQN, PPO, and SAC, the framework encompasses 192 attack–defense combinations. Systematic evaluations on LunarLander and Highway-v0 reveal substantial environment-dependent variations in agent robustness, demonstrate that certain defense methods can be detrimental, and consistently identify temporal smoothing as an effective strategy for enhancing robustness. The code is publicly released.
0 citationsRead paper