Roadside perception has a wider sensing range than onboard detection, providing enhanced sensory information for intelligent transportation systems, thus gaining increasing attention in recent years. However, directly applying onboard perception algorithms on roadside detectors (RSD) is infeasible due to the challenging requirements for high maneuvering target tracking and discriminating highly similar targets. Therefore, an Interactive Siamese Network (ISN) is proposed in this paper to overcome the roadside perception difficulties. Specifically, an interactive similarity contrast encoder-decoder has been developed within the ISN tracker. The sensitivity of algorithm to rapid changes in vehicle states is enhanced through the adaptive adjustment of weights assigned to critical tracking parameters within the loss function. This strategy enhances the ISN tracking effect for long-term trajectories for high maneuvering driving. Then, a global trajectory optimization unit is integrated into the ISN tracker. A trajectory similarity threshold is established to conduct cross-association analysis on similar trajectories, followed by iterative operations on adjacent trajectories. This approach further enhances the resolution of similar trajectories while ensuring the optimality of global trajectories. The proposed method is evaluated on the DAIR-V2X dataset and compared against current state-of-the-art methods. The experimental results verify that the proposed method provides efficient and accurate estimations and tracks vehicles in the intersection area, affording better accuracy and recall rate in high maneuvering target tracking than existing methods.

An accurate longitudinal dynamics model is essential for state estimation and control of autonomous vehicles. However, existing physical models suffer from limited working conditions and large errors, while pure data-driven models require massive amounts of driving data to cover working conditions fully. To address these issues, a hybrid architecture composed of a neural network-based traction model, a recursive least square-based parameter estimator, and a physics-based dynamics model is proposed for longitudinal dynamics modeling, in which the parameter estimator is used for mass and modified rolling friction coefficient estimation. Under this architecture, the longitudinal dynamics model can be established using limited driving data collected on a test field with a given load, and achieve precise vehicle dynamics characterization under various roads and loads. To design the neural network for traction description, the dynamics of vehicle powertrain and braking systems are analyzed, and a physics-guided neural network, which fully considers the traction transmission characteristics, is formulated. For model training, a two-stage hybrid model training method is proposed, which can train the hybrid model with the co-existence of unknown network and physical parameters. Results demonstrate that the proposed hybrid model can realize accurate parameter estimation and vehicle longitudinal dynamics modeling using limited driving data collected at a test field under a given load, especially with excellent generalization performance under different loads and roads.

Collision risk modeling with multiple surrounding target vehicles (TVs) is essential for host vehicle (HV) trajectory planning, especially considering challenging TV lateral behaviors. Existing motion-compensated spatial methods ignore TV lateral motion states such as lateral velocity and yaw rate, so that TV lateral behavior cannot be described accurately. Aiming at high-accuracy collision risk modeling, a twisted Gaussian risk model using both longitudinal and lateral motion states for TV behavior description is proposed. Firstly, the HV-TVs system is treated as the superposition of multiple HV-TV units, and a Gaussian risk model is adopted for the collision risk description of the HV-TV unit. Then, by expanding the variances, TV longitudinal and lateral velocities are considered. At last, a twisted Gaussian risk model considering TV yaw rate is constructed based on the projection of the Gaussian risk model. With this twisted Gaussian risk model, TV longitudinal-lateral motion states are considered simultaneously, and TV behaviors can be described for HV-TVs collision risk modeling. For HV trajectory planning, trajectory candidates generated by the maneuver-inspired method are evaluated via the proposed risk model, and the safe and efficient trajectory is selected. Simulation and hardware-in-the-loop experimental results show that the proposed method considering TV longitudinal-lateral motion states allows HV to operate more safely and efficiently than the conventional method.

Accurate trajectory prediction of surrounding vehicles is vital for automated vehicles to achieve high-level driving safety in complex situations. However, most state-of-the-art approaches for multi-vehicle trajectory prediction ignore vehicle motion uncertainty caused by different driving styles. Moreover, the interrelationship between the vehicle and the environment is seldom considered. To address the above problems, this paper proposes a driving risk map-integrated deep learning (DRM-DL) method for interactive trajectory prediction of surrounding vehicles, which comprehensively considers the motion uncertainty, trajectory intention uncertainty and interactions among vehicles, lane lines and road boundaries. Specifically, we adopt a conditional variational autoencoder (CVAE) to generate the candidate trajectories, in which the motion uncertainty is considered using a conditional Gaussian distribution. Furthermore, a driving risk map is constructed to realize a unified and interpretable representation of vehicle-vehicle and vehicle-environment interactions. The probability of each candidate trajectory is assigned using a trajectory probability model and a random selection is adopted to select a guided trajectory, which simulates the driver’s trajectory intention uncertainty. Finally, a relearning module is designed to obtain the precise trajectory prediction for surrounding vehicles. The proposed method is evaluated on the HighD dataset, and the results demonstrate a more accurate and reliable trajectory prediction for surrounding vehicles compared with state-of-the-art methods.

A timely understanding of target vehicles (TVs) lateral behavior is essential for the decision-making and control of host vehicle. Existing physical model-based methods such as motion-based method and multiple centerline-based method are generally constructed based on TV pose and longitudinal velocity, and tend to ignore TV preview driving characteristic and other useful information such as lateral velocity and yaw rate. To address these issues, a driver preview and multiple centerline model-based probabilistic behavior recognition architecture is proposed for timely and accurate TV lateral behavior prediction. Firstly, a driver preview model is used to describe vehicle preview driving characteristic, and TV preview lateral offset and preview lateral velocity are calculated with TV states and road reference information. Then, the preview lateral offset and preview lateral velocity are combined with multiple centerline model for TV lateral behavior reasoning based on the interacting multiple model-based probabilistic behavior recognition algorithm. With this method, TV preview driving characteristic and lateral motion states are combined for precise TV lateral behavior description. Furthermore, to predict short-term lateral behavior, a preview lateral velocity-dependent transition probability matrix model constructed with Gaussian cumulative distribution function is proposed. Simulation and experimental results show that the proposed method considering vehicle preview driving characteristic predicts TV lateral behavior earlier than the conventional method.

Preceding target vehicle (PTV) motion recognition play a pivotal role in autonomous vehicles. Motion states such as yaw rate, longitudinal and lateral velocity are critical for ego vehicle decision-making and control. However, lateral states of a PTV can hardly be measured directly by common onboard sensors and the PTV lateral state estimation has been seldom addressed in existing literatures. In this paper, a novel estimation scheme based on multiple neural network ensemble is proposed for PTV lateral state estimation. First, PTV lateral kinematics is presented based on vehicle-road relationship and a novel PTV lateral motion model is constructed to interpret the PTV lateral motion. Then, neural network observer with the PTV lateral kinematics as prior knowledge is designed and training data are collected in simulation environment. The neural network observer is trained using Levenberg-Marquardt backpropagation with Bayesian regularization (LMBR) to improve the generalization capability. Finally, to further improve the performance of the neural network estimation method, multiple neural network observers are integrated by weighted averaging strategy. The effectiveness of proposed approach is verified through hardware-in-the-Ioop (HiL) experiments conducted in designed verification scenarios, and compared with model-based method and other three learning methods. The experiment results reveal that the proposed method outperforms other typical methods and achieves accurate estimation of the PTV lateral states.

Automated vehicle control requires full knowledge of motion behavior of the preceding target vehicles (PTVs), and the states such as longitudinal/lateral velocity and yaw rate are critical for the PTV behavior description. However, the PTV's lateral states estimation have seldom been addressed in the state-of-the-art literatures. Aimed at providing reliable PTV lateral states, this paper presents a novel combined model-based estimation scheme. Different from the conventional PTV models, the proposed model is constructed based on the host-target vehicle dynamics and road constraints. Specifically, steering angle of the PTV is included in the state vector. The measurements, such as heading angle, road curvature, and lateral distance to the lane center, are available from an onboard vision system. As a vision system inevitably has measurement delay, a modified Kalman filter is developed to address the sampling issue. To verify the proposed approach, hardware-in-the-loop experiments are conducted in designed testing scenarios.