Accurately estimating the real-time state of dynamic systems from noisy measurements is a fundamental task in signal processing. State-space models (SSM) are commonly utilized to model system dynamics and account for uncertainty in both system state evolution and measurement data. Kalman filter (KF) and its variants are widely recognized for their low computational complexity and ability to address state evolution and measurement update. In practice, accurately modeling the system dynamics and obtaining model parameters are difficult, especially for nonlinear systems. These tasks often require predefined motion and measurement models that rely on coarse approximations and extensive manual tuning. In this paper, we propose DL-KF, a robust Deep Learning aided Kalman Filter designed to perform KF in nonlinear dynamic environments where the process model is unavailable and the system noise is also unknown. However, a linear measurement model is assumed. DL-KF utilizes a Long Short-Term Memory (LSTM) network to capture nonlinear process dynamics and generate state priors from historical measurements. Additionally, a Gated Recurrent Unit (GRU) is utilized to independently learn innovation covariance, addressing measurement noise uncertainty, followed by the computation of the Kalman gain within the Kalman filter framework. By integrating an SSM framework with deep learning modules, the proposed method strikes the balance between the flexibility of data-driven approach and the interpretability of classical Kalman based approaches. Empirical results demonstrate that DL-KF significantly outperforms traditional model-based filters, including the KF, EKF, and UKF. Furthermore, it surpasses state-of-the-art hybrid methods such as KalmanNet, Split-KalmanNet, and AKNet, particularly in challenging scenarios involving high nonlinearity (e.g., the Lorenz attractor) and maneuver-induced model mismatches. The algorithm’s efficacy and robustness are further validated through extensive experiments on real-world tunnel radar and NCLT datasets.

Tracking maneuvering targets, such as connected and automated vehicles, requires modeling their movements with pre-defined kinematic models. However, sudden and unpredictable maneuvers often lead to model mismatch, resulting in significant peak tracking errors. To address this problem, we propose a maneuver detection-aided deep learning multiple model filter (MD-DL-MM) technique for target tracking and state estimation, designed to suppress peak errors and improve tracking performance. The core innovation of the MD-DL-MM technique is the integration of a self-attention-based discrimination network, which dynamically determines the weights of the target’s motion models. To consider the potential impact of a specific kinematic models on the state estimation accuracy, a statistical hypothesis testing-based method is introduced to evaluate the validity of kinematic models. This metric system-atically examines the suitability of the motion model, ensuring that the most appropriate model is selected and applied at each stage of the tracking process. On that basis, two distinct state estimation methods are designed. Specifically, when the kinematic model is more appropriate and the target exhibits non-maneuvering, state estimates are obtained using a recursive model-based Kalman filter (MB-KF), which provides optimal estimation with the minimum mean square error (MMSE). On the other hand, when the target exhibits sudden maneuvers or unpredictable maneuvers, a data-driven learning-based network is utilized to achieve high-precision state estimation. Extensive simulation results and real-world experimental data demonstrate that the proposed algorithm outperforms traditional methods such as the Singer, current statistical (CS), and the interactive multiple model (IMM) algorithm, as well as deep learning-based algorithms like DeepMTT and KalmanNet. The proposed algorithm achieves superior performance in terms of stability, computational efficiency, and tracking accuracy across diverse scenarios.

The integration of the global navigation satellite system (GNSS), odometer, and inertial navigation system (INS) holds significant potentials for achieving high-precision vehicle localization. However, GNSS is vulnerable to obstructions and jamming, and the odometer is unreliable in harsh road conditions. These factors can lead to cumulative positioning errors in GNSS-denied environments. To address these issues, a novel multi-source information fusion based vehicle localization method that integrates an onboard binocular camera, an INS, and smart road studs—Internet of Things (IoT) devices extensively used for road safety and data collection in intelligent transportation systems— is introduced. We construct a position measurement model directly in the camera coordinate system through an enhanced You Only Look Once 8th version (YOLOv8) algorithm for smart road stud detection, combined with binocular vision measurement and position transformations. Additionally, we propose a method to enhance vehicle localization accuracy by integrating vehicle speed without relying on additional hardware speed sensors. The vehicle’s speed is estimated from the image sequences captured by the onboard camera using a deep neural network (DNN), named Speed-Net. The final navigation results are produced by fusing the smart road stud aided positioning information, the estimated vehicle speed, and INS data through an error-state extended Kalman filter (ESEKF). Real-world experiments demonstrate the effectiveness of the proposed Smart road stud (SRS)/Velocity/INS integrated vehicle localization method.

Traditional methods for tracking maneuvering targets, such as connected and automated vehicles, face significant challenges in complex driving environments due to the need for constant adjustments to the state transition model to match the target’s motion. These adjustments often result in decision-making delays and competition between models. Furthermore, the widely adopted first-order Markov assumption frequently fails to capture time-dependent motion patterns, leading to information loss. Although the Interacting Multiple Model (IMM) algorithm mitigates some of these issues by employing multiple motion models, it still struggles with accurately identifying motion patterns, delays in maneuver detection and reduced tracking accuracy. To address these problems, we propose a novel approach that combines deep neural networks with traditional IMM tracking methods. Leveraging the strength of deep learning in classification tasks, we introduce an attention mechanism to enhance motion model recognition. This leads to the development of an enhanced version of IMM, termed Attention-IMM. We evaluate our method on the widely-used LAST dataset and real word automated vehicles data. The results demonstrate that Attention-IMM achieves superior performance in both tracking accuracy and the timeliness and accuracy of model decision-making, offering a robust and efficient solution for maneuvering target tracking.

Sensor fusion plays an increasingly important role in real-time traffic perception using roadside sensing devices because the use of single type of sensors often fail to deliver satisfactory performance in certain harsh environment. This paper investigates asynchronous data fusion for real-time vehicle tracking with inaccurate and randomly delayed measurements from millimeter-wave (MMW) radars and magnetic sensors in tunnel environment. We first propose a multisensor data association algorithm to assign the measurements of MMW radar and magnetic sensors to a particular vehicle. A tracking algorithm is then designed to asynchronously update the current vehicle states with randomly delayed magnetic sensor measurements. The proposed algorithm is implemented in the Xianfengding Tunnel, Jiangxi Province, China. Experiments validate the proposed method's accuracy using real data. The method and collected data form the basis of a real-time digital twin system to support advanced traffic management. The fusion results and measurement dataset are available at https://github.com/futianxuan/data.

Sensor fusion plays an increasingly important role in real-time traffic perception using roadside sensing devices because the use of single type of sensors often fail to deliver satisfactory performance in certain harsh environment. This paper investigates asynchronous data fusion for lane-level vehicle tracking with randomly delayed measurements and inaccurate detections from millimeter wave (MMW) radars and magnetic sensors in tunnels, where vehicle tracking with single type of sensors can not meet the requirements of reliable and accurate lane-level tracking due to inaccurate radar detections at far distances, noisy radar detections in tunnel environment, and missed or false vehicle detections by magnetic sensors. A multisensor data association algorithm is first designed to assign the measurements of MMW radar and magnetic sensors to a particular vehicle. A multi-lane estimation model is then developed, which employs Bayesian weight mixture filtering to fuse MMW radar and magnetic sensor measurements and to estimate the lane in which a vehicle is located. Finally, the proposed algorithm is implemented in a real environment - the Xianfengding Tunnel located in Jiangxi Province, China. Experiments are conducted to validate the accuracy of the proposed method using real data. The proposed method and the collected data are further integrated to establish a real-time digital twin system aimed at supporting advanced traffic management. The fusion results and the real radar measurement dataset of the tunnel are made available athttps://github.com/futianxuan/data.

Smart road studs have been extensively deployed as road safety and data collection devices. Accurate and reliable detection of smart road studs and its further integration into the perception and control modules of connected and autonomous vehicles (CAVs) undoubtedly benefit road boundary detection, localization of CAVs and augument the safety of CAVs’ driving. This work investigates real-time, accurate and reliable detection of smart road studs, which is a challenging task for CAVs because existing methods fail to achieve accurate and real-time smart road stud detection, especially in harsh road environment. To address these challenges, we first build a real-world smart road stud dataset, and then propose and validate a lightweight and efficient smart road stud detection model based on the you only look once 8th version (YOLOv8), called SRS-YOLO. First, a Squeeze-and-Excitation (SE) attention module is used to improve the coarse-to-fine (C2F) module to differentiate the channel importance of feature maps and improve the detection accuracy of smart road studs. Second, a novel downsampling module (DownS) that integrates the average pooling and the max pooling is designed to reduce the number of parameters and minimize information loss during the downsampling process. Third, the loss function is replaced with the Normalized Wasserstein Distance (NWD) loss to alleviate the sensitivity to location deviations when computing the loss for small targets. The experimental results demonstrate that the proposed SRS-YOLO outperforms other state-of-the-art methods, and achieves a 87.92% mean average precision at a real-time speed of 78 frames/s. Our dataset is available at:https://github.com/wky-xidian/smart-road-stud-dataset.

Integration of global navigation satellite system (GNSS) and inertial navigation system (INS) presents significant potential for high-precision vehicle localization. However, this approach suffers from cumulative INS errors in GNSS-denied environments. To address this issue, this paper proposes a method to correct the navigation errors due to INS by using Internet of Things (IoT)-based vision positioning. More specifically, this method employs a binocular camera to assist in obtaining the vision positioning of the vehicle through recognition of LED lights installed in a type of ubiquitously deployed IoT devices that are increasingly used in smart transportation systems. The final navigation outcomes are attained by fusing the vision positioning results with INS information using an unscented Kalman filter (UKF). Real-world experiment results validate the effectiveness of the proposed vehicle localization method.