Zhiqun Hu

Achieving accurate and flexible localization in global navigation satellite system (GNSS)-denied underground spaces is critical for Internet of Things (IoT)-enabled logistics and personnel operations. To this end, this paper proposes a scalable multimodal localization framework requiring only a single ultra-wideband (UWB) transmitter connected to a distributed antenna system, avoiding the deployment of multiple additional anchors. An adjacency-masked, change-point-aware hidden Markov model (AC-HMM) is developed for region identification using only two-path UWB measurements, which avoids full channel impulse response (CIR) processing and reduces computational complexity. A multi-scale factor-graph maximum a posteriori inference method (MS-FGM) is then proposed for dynamic localization by fusing UWB and magnetic-field residuals with region and motion constraint factors. Multi-scale temporal aggregation is further introduced to mitigate motion-induced fluctuations and improve localization accuracy and stability. Experiments conducted along the roadways of an underground parking lot demonstrate an average region classification accuracy of 96.81% and a mean positioning error of 0.88 m, outperforming existing methods by up to 42.19%.

Vision foundation models (VFMs) enable high-quality 2D instance masks, yet their lifted pseudo-point clouds suffer from scale ambiguity, structural noise, and temporal inconsistency, limiting their utility in 3D annotation. Existing automatic labeling methods either rely on expensive light detection and ranging (LiDAR) sensors or fail to enforce physical plausibility in dynamic roadside scenes. This study proposes a LiDAR-free radar–visual auto-labeling framework that leverages cross-modal spatio-temporal consistency between millimeter-wave radar trajectories and visual pseudo-point clouds to self-correct 3D geometry. The method first associates radar points, 2D masks, and pseudo-point clouds into object-centric sequences. Then, an uncertainty-aware pose fusion module combines motion-derived and structure-derived orientations using automatically solved road priors. Finally, the pseudo-point cloud is refined in canonical space by optimizing stable semantic landmarks from temporally consistent masks and propagating their corrections globally. Evaluated on a real-world roadside dataset, the method achieves 49.1% bird’s-eye-view (BEV) intersection over union (IoU) and 43.0% 3D IoU, outperforming a radar–camera fusion baseline by 5.5/5.9 points. Downstream experiments further show that the generated pseudo-labels and semantic enhancement are useful under the evaluated detector configurations, while broader validation remains future work.

Deep reinforcement learning (DRL)-based traffic signal control (TSC) algorithms often suffer from overfitting to static training environments and perform poorly in unseen traffic scenarios. Two major challenges remain in previous methods: the scarcity and homogeneity of training data, and the limited adaptability to harsh or challenging traffic environments. To address these issues, we propose RAW-MetaRL, a Robust Adversarial Weighted Meta-Reinforcement Learning framework. RAW-MetaRL introduces an adversarial environment agent that adaptively generates increasingly challenging and diverse traffic environments based on the current policy’s performance, where synthetic environments are optimized to expose weaknesses in the current control policy. Furthermore, we design a weighted meta-learning framework that alternates between local-level adaptation on individual tasks and global-level adaptation across a sampled set of tasks, aiming to effectively train the meta-agent. A meta weight generator is incorporated to prioritize rare and critical environments, enabling the meta agent to generalize effectively across diverse and previously unseen environments. Extensive experiments on real-world and synthetic datasets demonstrate that RAW-MetaRL significantly outperforms existing methods in adaptability and performance across diverse traffic environments.

With the advancement of artificial intelligence, Deep Reinforcement Learning (DRL) has been widely applied to traffic signal control (TSC). However, DRL-based TSC methods often require extensive training samples and struggle to generalize across heterogeneous traffic environments, such as different intersection types and traffic flows. Retraining the DRL agent for new environments inevitably increases both costs and time investments, limiting the flexibility of these intelligent TSC algorithms in practical applications. In this paper, we propose an inductive meta-DRL-based traffic signal control method, named IM-TSC, for heterogeneous environments to address these issues. First, we design a feature extraction module composed of a self-feature induction network and a neighborhood feature extraction network to obtain accurate and comprehensive traffic representations, laying the foundation for generalization and coordination. Furthermore, we decouple the scenario inference and agent training processes and design a scenario inference network to obtain contextual encoding for the current intersection scenario through contrastive learning, thereby promoting the transfer of meta-knowledge across similar scenarios. Finally, we design a meta-training framework for the DRL agent with a continuous action space, enabling efficient training of the meta-learner by alternating between local-level learning and global-level learning. We evaluate the proposed IM-TSC method across multiple heterogeneous traffic environments. Experimental results demonstrate that the meta-learner trained with the IM-TSC method can adapt to heterogeneous environments more quickly and effectively. Compared to the suboptimal baseline, it achieves average reductions of 9.20% in the number of stops, 11.55% in waiting time, and 8.29% in fuel consumption.

Roadside perception is a key component in intelligent transportation systems. In this paper, we present a novel three-dimensional (3D) extended object tracking (EOT) method, which simultaneously estimates the object kinematics and extent state, in roadside perception using both the radar and camera data. Because of the influence of sensor viewing angle and limited angle resolution, radar measurements from objects are sparse and non-uniformly distributed, leading to inaccuracies in object extent and position estimation. To address this problem, we present a novel spherical Gaussian function weighted Gaussian mixture model. This model assumes that radar measurements originate from a series of probabilistic weighted radar reflectors on the vehicle’s extent. Additionally, we utilize visual detection of vehicle keypoints to provide additional information on the positions of radar reflectors. Since keypoints may not always correspond to radar reflectors, we propose an elastic skeleton fusion mechanism, which constructs a virtual force to establish the relationship between the radar reflectors on the vehicle and its extent. Furthermore, to better describe the kinematic state of the vehicle and constrain its extent state, we develop a new 3D constant turn rate and velocity motion model, considering the complex 3D motion of the vehicle relative to the roadside sensor. Finally, we apply variational Bayesian approximation to the intractable measurement update step to enable recursive Bayesian estimation of the object’s state. Simulation results using Carla simulator and experimental results on the nuScenes dataset demonstrate the effectiveness and superiority of the proposed method in comparison to several state-of-the-art 3D EOT methods.

The integration of cameras and radar has gained significant attention in the context of roadside perception. However, existing calibration techniques face challenges when applied to roadside infrastructures due to unstable feature acquisition and large computational overhead. In this paper, we propose an automatic calibration framework for camera-radar extrinsic calibration. The core of this framework lies in leveraging trajectory events from sensor measurement to eliminate modality disparity and achieve stable target representation at a low cost. By decomposing single events into multiple base trajectory behaviors, the extracted radar and camera events are associated in a multistage event association pipeline. Moreover, we use the proposed calibration framework to construct an automatic calibration system that can be integrated into prevalent roadside perception systems. The calibration system continuously monitors the spatial relationship between sensors and quickly eliminates any possible de-calibrations, which greatly reduces re-calibration costs. The experiments conducted on real-world scenarios in Beijing North Third Ring Middle Road have demonstrated the feasibility and quick response of the system, as well as an accuracy leap that surpasses other methods in camera-radar calibration.

Owing to the inherent advantages of high flexibility and strong Line-of-Sight (LoS) links, as well as the continuous power supply, tethered Unmanned aerial vehicle (TUAV) connected to ground charging stations (GCSs) is regarded as a feasible and effective solution to support stable emergency communications. One of the key issues in the TUAV network deployment is to ensure the flight safety considering the potential risk of tether tangling as well as collision along with TUAVs’ movements. In this paper, we propose a theoretical method for analyzing mobile TUAV networks with spacial repulsion constrain. Under the nearest association criterion, we apply the Matérn hard-core point process (MHCPP) to derive the steady-state spacial distance distribution between the TUAV and the reference ground user equipment (GUE). In the vertical direction, we assume that all TUAVs dynamically adjust their altitude following a random waypoint (RWP) mobility model. Then, closed-form expressions for the network performance metrics such as coverage probability and handover probability are presented, jointly considering the path loss model and generalized Nakagami-m fading channels. Our analytical results, validated through Monte Carlo simulations, demonstrate the efficiency and accuracy of the proposed method.

Multiple object tracking is a key component of intelligent transportation systems. Most existing tracking methods rely on deep learning frameworks, requiring extensive high-quality three-dimensional (3D) annotated datasets, which limits their deployment in real-world scenarios. To address this problem, we present a novel 3D multiple extended object tracking (MEOT) method by fusing radar and camera data with only pixel annotations. The fusion framework employs a joint detection and tracking (JDT) strategy based on random finite set (RFS) theory, and we propose a multimodal trajectory Poisson multi-Bernoulli (MTPMB) tracker to incorporate multimodal data from radar and camera for the first time. In the update stage, we propose a novel back-projection method and its corresponding visual measurement model based on a mixture density network (MDN) that automatically learns the transformation from pixel coordinates to 3D road surface, allowing visual measurements to be accurately back-projected to 3D space. Additionally, we introduce a doped association method that uses visual measurements to assist in splitting radar measurements from nearby objects and establish a correlation between radar and visual measurements. Experimental results demonstrate the effectiveness and superiority of the proposed method compared to several state-of-the-art 3D MEOT methods without 3D annotations. Code is available at https://github.com/RadarCameraFusionTeam-BUPT/ES-MEOT.