Safe and executable trajectory planning in urban environments requires jointly considering traffic-related elements, ego behavior, and vehicle kinematics, posing challenges to the real-time performance and convergence of optimization-based methods. This paper proposes a behavior-guided optimization framework that structures the solution space at the spatiotemporal drivable domain level to facilitate fast and stable convergence. The planning space is partitioned into modular spatiotemporal domains, termed Behavior Cells (BCs), which encode ego motion feasibility and traffic-induced decisions. Feasible high-level behaviors are systematically enumerated through structured BC combinations and evaluated via a finite-horizon Markov decision process. The selected BC combination defines a continuous, behavior-consistent solution space, within which a dynamic two-stage optimization progressively restores the full planning formulation, enabling efficient and robust trajectory generation. Extensive simulations across diverse traffic scenarios demonstrate consistent reliability and real-time performance under varying traffic densities. On-road experiments further validate effectiveness in real-world urban environments. More detailed results are available at:https://lshasd123.github.io/Behavior-Cells/

Deep learning (DL) has shown state-of-the-art performance in trajectory prediction, which is critical to safe navigation in autonomous driving (AD). However, most DL-based methods suffer from catastrophic forgetting, where adapting to a new distribution may cause significant performance degradation in previously learned ones. Such inability to retain learned knowledge limits their applicability in the real world, where AD systems need to operate across varying scenarios with dynamic distributions. As revealed by neuroscience, the hippocampal circuit plays a crucial role in memory replay, effectively reconstructing learned knowledge based on limited resources. Inspired by this, we propose a hippocampal circuit-inspired continual learning method (H2C) for trajectory prediction across varying scenarios. H2C retains prior knowledge by selectively recalling a small subset of learned samples. First, two complementary strategies are developed to select the subset to represent learned knowledge. Specifically, one strategy maximizes inter-sample diversity to represent the distinctive knowledge, and the other estimates the overall knowledge by equiprobable sampling. Then, H2C updates via a memory replay loss function calculated by these selected samples to retain knowledge while learning new data. Experiments based on various scenarios from the INTERACTION dataset are designed to evaluate H2C. Experimental results show that H2C reduces catastrophic forgetting of DL baselines by 22.71% on average in a task-free manner, without relying on manually informed distributional shifts. The implementation is available at https://github.com/BIT-Jack/H2C-lifelong.

Maps play an important role in autonomous driving systems. The recently proposed 3D Gaussian Splatting (3D-GS) produces rendering-quality explicit scene reconstruction results, demonstrating the potential for map construction in autonomous driving scenarios. However, because of the time and computational costs involved in generating Gaussian scenes, how to update the map becomes a significant challenge. In this paper, we propose LT-Gaussian, a map update method for 3D-GS-based maps. LT-Gaussian consists of three main components: Multimodal Gaussian Splatting, Structural Change Detection Module, and Gaussian-Map Update Module. Firstly, the Gaussian map of the old scene is generated using our proposed Multimodal Gaussian Splatting. Subsequently, during the map update process, we compare the outdated Gaussian map with the current LiDAR data stream to identify structural changes. Finally, we perform targeted updates to the Gaussian-map to generate an up-to-date map. We establish a benchmark for map updating on the nuScenes dataset to quantitatively evaluate our method. The experimental results show that LT-Gaussian can effectively and efficiently update the Gaussian-map, handling common environmental changes in autonomous driving scenarios. Furthermore, by taking full advantage of information from both new and old scenes, LT-Gaussian is able to produce higher quality reconstruction results compared to map update strategies that reconstruct maps from scratch. Our open-source code is available at https://github.com/ChengLuqi/LT-gaussian.

Traffic knowledge is essential for safe, efficient, and regulation-compliant autonomous driving. While existing path planning methods often encode only partial traffic rules or focus on specific scenarios, they lack adaptability to diverse and dynamic environments. The Internet of Things (IoT) enables vehicles to access rich, real-time traffic knowledge, yet most IoT-based approaches emphasize long-distance route optimization and overlook local path planning. This paper proposes a traffic knowledge-augmented path planning model (KARS) that integrates a knowledge graph representation of traffic knowledge with a reachable set-based planning framework in IoT environments. The knowledge graph systematically encodes spatial, temporal, and speed constraints, while the reachable set method enables the planner to dynamically adapt to real-time traffic knowledge. KARS is evaluated in multi-constraint simulation environments with static, semi-dynamic, and dynamic traffic knowledge settings. Compared to representative baselines, it improves driving safety by increasing traffic success rates in complex environments, raises maximum driving speed to enhance operational efficiency, and produces smoother acceleration profiles for improved ride comfort. These results validate the effectiveness of combining traffic knowledge with reachable set planning for safe, efficient, and adaptable autonomous driving in IoT-enabled urban scenarios.

Risk assessment plays an essential role in the improvement of driving safety for intelligent vehicles. Current methods ignoring the predictive and personalized impact of driving policies weaken the effectiveness of risk assessment and lead to human-machine conflicts. By combining subjective cognition of drivers and objective risk metrics, a policy-oriented cognitive risk map (POCRM) is proposed in this paper to encode different driving policies in risk assessment for lane-changing scenarios. To obtain the objective safety metrics, insecurity quantification is built based on the fuzzy theory and fault tree analysis. The subjective cognition of drivers for different driving policies is modeled by deep successor representation and encoded in POCRM using deep reinforcement learning. Driving data collected from the public dataset for realistic traffic environment are used to evaluate the proposed POCRM. The experimental results show that the risk map can take into account future risks and provide driving advice that balances human-machine conflicts with safety in scenarios where drivers can or cannot correctly perceive risk.

Accurate assessment of the cyclist risk is a crucial task for the safety system of autonomous vehicles (AVs). This paper proposes a framework for defining and evaluating cyclist risk levels, considering behavioral cues. The framework comprises three modules: the cyclist graph construction (CGC) module, the risk label generation (RLG) module, and the risk assessment (RA) module. The CGC module constructs a spatiotemporal graph model of the cyclist with both the behavioral and risk information. The RLG module leverages the graph representation method (GRM) to extract features and assigns risk labels using unsupervised learning. The RA module employs spatiotemporal graph convolutional networks (ST-GCN) to extract features from the cyclist graph. Additionally, it facilitates feature fusion through interactions between the human body and the two-wheeler and between hierarchical levels. The fused features, along with the risk labels, are used to train a classifier for the risk assessment of cyclists. The proposed framework is validated using real-world data, and the comparative results with state-of-the-art methods demonstrate the effectiveness and accuracy of the proposed approach in cyclist risk assessment in mixed traffic.

Trajectory prediction is a fundamental problem and challenge for autonomous vehicles. Early works mainly focused on designing complicated architectures for deep-learning-based prediction models in normal-illumination environments, which fail in dealing with low-light conditions. The paper proposes a novel approach for trajectory prediction in low-illumination scenarios by leveraging multi-stream information fusion, which integrates image, optical flow, and object trajectory information. This is achieved by applying Convolutional Neural Network-based (CNN) Long Short-term Memory (LSTM) networks to extract temporal information from the image channel, Spatial-Temporal Graph Convolutional Network (ST-GCN) to model relative motion between adjacent camera frames through the optical flow channel, and recognizing high-level interactions between vehicles in the trajectory channel. Further, to investigate the reliability of the model in low-illumination scenarios, epistemic uncertainty estimation is conducted by applying Monte Carlo Dropout. The proposed approach is validated on HEV-I and newly generated Dark-HEV-I datasets focusing on graph-based interaction understanding and low illumination conditions. The experimental results show improved performance compared to baselines in both standard and low-illumination scenarios. Importantly, our approach is generic and applicable to scenarios with different types of perception data. The source code is available at https://github.com/TommyGong08/MSIF.

The deep understanding of the behaviors of traffic participants is essential to guarantee the safety of automated vehicles (AV) in mixed traffic with vulnerable road users (VRUs). Precise trajectory prediction of traffic participants can provide reasonable solution space for motion planning of AV. Early works mainly focused on handcrafting the feature representation and designing complicated architectures in deep learning-based prediction models. However, these approaches overlooked the fact that different road users perceive the safety of the same interaction differently and also exhibit heterogeneous risk-taking styles. In this paper, we will develop a model for trajectory prediction based on risk-taking styles. The model accounts for the expected positions and occupancy of traffic participants in the surrounding environment. It consists of two sequential steps: risk-taking styles of multi-modal road users under interactive scenes are first clustered, and then reformulated in the heterogeneous graph model for trajectory prediction. The model is validated by the driving data collected on the urban road using a public dataset. Comparative experiments demonstrate that the proposed method can predict the trajectory of traffic participants much more accurately than the state-of-the-art methods.

Developing autonomous vehicles (AVs) helps improve the road safety and traffic efficiency of intelligent transportation systems (ITS). Accurately predicting the trajectories of traffic participants is essential to the decision-making and motion planning of AVs in interactive scenarios. Recently, learning-based trajectory predictors have shown state-of-the-art performance in highway or urban areas. However, most existing learning-based models trained with fixed datasets may perform poorly in continuously changing scenarios. Specifically, they may not perform well in learned scenarios after learning the new one. This phenomenon is called “catastrophic forgetting”. Few studies investigate trajectory predictions in continuous scenarios, where catastrophic forgetting may happen. To handle this problem, first, a novel continual learning (CL) approach for vehicle trajectory prediction is proposed in this paper. Then, inspired by brain science, a dynamic memory mechanism is developed by utilizing the measurement of traffic divergence between scenarios, which balances the performance and training efficiency of the proposed CL approach. Finally, datasets collected from different locations are used to design continual training and testing methods in experiments. Experimental results show that the proposed approach achieves consistently high prediction accuracy in continuous scenarios without re-training, which mitigates catastrophic forgetting compared to non-CL approaches. The implementation of the proposed approach is publicly available athttps://github.com/BIT-Jack/D-GSM.

Accurate recognition of driver behaviours is the basis for a reliable driver assistance system. This paper proposes a novel fusion framework for driver behaviour recognition that utilises the traffic scene and driver gaze information. The proposed framework is based on the graph neural network (GNN) and contains three modules, namely, the gaze analysing (GA) module, scene understanding (SU) module and the information fusion (IF) module. The GA module is used to obtain gaze images of drivers, and extract the gaze features from the images. The SU module provides trajectory predictions for surrounding vehicles, motorcycles, bicycles and other traffic participants. The GA and SU modules are parallel and the outputs of both modules are sent to the IF module that fuses the gaze and scene information using the attention mechanism and recognises the driving behaviours through a combined classifier. The proposed framework is verified on a naturalistic driving dataset. The comparative experiments with the state-of-the-art methods demonstrate that the proposed framework has superior performance for driving behaviour recognition in various situations.

Precisely modeling interactions and accurately predicting trajectories of surrounding vehicles are essential to the decision-making and path-planning of intelligent vehicles. This paper proposes a novel framework based on ensemble learning to improve the performance of trajectory predictions in interactive scenarios. The framework is termed Interactive Ensemble Trajectory Predictor (IETP). IETP assembles interaction-aware trajectory predictors as base learners to build an ensemble learner. Firstly, each base learner in IETP observes historical trajectories of vehicles in the scene. Then each base learner handles interactions between vehicles to predict trajectories. Finally, an ensemble learner is built to predict trajectories by applying two ensemble strategies on the predictions from all base learners. Predictions generated by the ensemble learner are final outputs of IETP. In this study, three experiments using different data are conducted based on the NGSIM dataset. Experimental results show that IETP improves the predicting accuracy and decreases the variance of errors compared to base learners. In addition, IETP exceeds baseline models with 50% of the training data, indicating that IETP is data-efficient. Moreover, the implementation of IETP is publicly available at https://github.com/BIT-Jack/IETP.

The path planning of unmanned differential steering vehicles (UDSVs) in the off-road environment not only needs to consider the non-complete constraints of vehicles but also faces the challenges of complex off-road terrains and obstacles. In this paper, an integrated path planning system is proposed to handle the influence of kinematic vehicle model, off-road terrains and obstacles systematically for the path planning of UDSVs. To improve the planning efficiency, a Pre-planning is designed and carried out using the Voronoi diagram established in the 3D environment with obstacles. By combining the potential field functions (PFF) related to passable obstacles and 3D terrains, an integrated PFF is defined to represent the movement cost of UDSV in the nonlinear optimal control (NOC) problem. Based on the NOC, a channel path planning (CPP) problem is formulated to avoid the untraceable path caused by the traditional line path planning (LPP). Simulation results show that the proposed system can plan a feasible path fast with the constraints from vehicle kinematics, obstacle avoidance and off-road terrains.

In complex and dynamic urban traffic scenarios, the accurate prediction of trajectories of surrounding traffic participants (vehicles, pedestrians, etc) with interactive behaviours plays an important role in the navigation and the motion planning of the ego vehicle. In this paper, based on the graph neural network (GNN), we propose a hierarchical GNN framework to model interactions of heterogeneous traffic participants (vehicles, pedestrians and riders) combined with LSTM to predict their trajectories. The proposed framework consists of two modules with two GNNs for interactive events recognition (IER) and trajectory prediction (TP). The IER module is used to recognise interactive events between traffic participants and the ego vehicle. With the recognised results as the input, the TP module is built for interactive trajectory prediction. In addition, to realise the multi-step prediction, a long short-term memory network (LSTM) is combined with GNN in the TP module. The proposed hierarchical framework is verified by the naturalistic driving data collected from the urban traffic environment. Comparative results with state-of-the-art methods indicate that the hierarchical GNN framework obtains an outstanding performance in the recognition of interactive events and the prediction of interactive behaviours.

Driver model adaptation (DMA) plays an essential role for driving behaviour modelling when there is a lack of sufficient data for training the new model. A new data-driven DMA method is proposed in this paper to realise the instance-level knowledge transfer between individual drivers. Using the importance-weighted transfer learning (IWTL), the data collected from one driver (source driver) can be directly used to train the model of another driver (target driver). Under the framework of IWTL, the relationship between two different drivers can be modelled by the importance weight (IW). Two estimation methods Kullback-Leibler (KL) Divergence and least-squares (LS), are used to estimate IW for each data instance by modelling the importance-weight function as a radial basis function (RBF). Experiments based on the driving simulator and real vehicle are carried out to test the performance of TL for steering behaviour adaptation during the overtaking manoeuvre. The experimental results show that the TL method can transfer the knowledge observed from one driver to another when training the new driver model without sufficient data by keeping the modelling error at a low level.

This paper presents a novel hierarchical reinforcement learning (HRL) framework for automated overtaking. The proposed framework is developed based on the semi-Markov decision process (SMDP) and motion primitives (MPs) which can be applied to different overtaking phases. Unlike the high-level decision and low-level control which are usually independent with each other, the high-level decision making and low-level control are combined by defining MPs with different time intervals. As for the high-level decision making, a SMDP Q-learning algorithm is adopted to realize decision-making of MPs. Besides, a development method of MPs used in the low-level control of automated overtaking is proposed. The performance of the HRL framework is tested in the simulation environment built in a driving simulator called CARLA. The results show that the HRL framework can determine the optimal trajectory under different driving styles of the overtaken vehicle.

In this paper, a Pedestrian Risky Level Prediction (PRLP) model based on the data collected by a vehicle-mounted sensing system is proposed. The prediction of pedestrian risky level is realized using a data-driven method, which contains two steps, namely pedestrian trajectory prediction and risky level classification. For trajectory prediction, an LSTM network is built and trained to predict the trajectory of pedestrian in the onboard camera view. For risky level classification, to avoid manually selecting thresholds for different risky levels, a hybrid method integrating K-Means Clustering (KMC), Kernel Principal Components Analysis (KPCA) and Kernel Support Vector Machine (SVM) is applied. The pedestrian risky level is divided into four classes by KPCA-KMC, namely low risk, medium risk, high risk and super-high risk. Experimental results show the capability of PRLP model to predict the risky level of pedestrian, which could help the intelligent vehicle to assess the risk of collision with pedestrians, and provide safety warnings to both vehicle and pedestrian.

The autonomous vehicles need to cooperate with the nearby vehicles to ensure driving safety, however, it is challenging to plan and follow the desired trajectory considering the nearby vehicles. This article proposes a cooperative driving strategy for the connected vehicles by integrating vehicle velocity prediction, motion planning, and robust fuzzy path-following control. The system uncertainties are considered to enhance the cooperation between the autonomous vehicle and the nearby vehicle. With the driving information obtained from the connected vehicles technique, the recurrent neural network is used to predict the nearby vehicle velocity. A motion planner is developed to provide the reference trajectory considering the velocity prediction errors. Then, a robust fuzzy path-following controller is designed to track the planned trajectory. The CarSim simulations are conducted to validate the proposed cooperative driving strategy. The simulation results show that the autonomous vehicle can avoid collisions with the nearby vehicle by applying the proposed driving strategy in the overtaking and the lane-changing scenarios.

Driver model adaptation (DMA) provides a way to model the target driver when sufficient data are not available. Traditional DMA methods running at the model level are restricted by the specific model structures and cannot make full use of the historical data. In this paper, a novel DMA framework based on transfer learning (TL) is proposed to deal with the adaptation of driver models in lane-changing scenarios at the data level. Under the proposed DMA framework, a new TL approach named DTW-LPA that combines dynamic time warping (DTW) and local Procrustes analysis (LPA) is developed. Using the DTW, the relationship between the datasets for different drivers can be found automatically. Based on this relationship, the LPA can transfer the data in the historical dataset to the dataset of a newly-involved driver (target driver). In this way, sufficient data can be obtained for the target driver. After the data transferring process, a proper modeling method, such as the Gaussian mixture regression (GMR), can be applied to train the model for the target driver. Data collected from a driving simulator and realistic driving scenes are used to validate the proposed method in various experiments. Compared with the GMR-only and GMR-MAP methods, the DTW-LPA shows better performance on the model accuracy with much lower predicting errors in most cases.

Because of the high accuracy and low cost, learning-based methods have been widely used to model driver behaviors in various scenarios. However, the performance of learning-based methods depend heavily on the quantity and coverage of the driving data. When the new driver with insufficient data is considered, the accuracy of these methods cannot be guaranteed any more. To solve this problem, the balanced distribution adaptation (BDA) is used to build the new driver's decision making model in the lane change (LC) scenario. Meanwhile, a transfer learning (TL) based regression model, modified BDA (MBDA) is proposed to predict the driver's steering behavior during the LC maneuver. Cross validation (CV) based model selection (MS) method is developed to obtain the optimal parameters in model training process. A series of experiments are carried out based on the simulated and naturalistic driving data to verify the TL based classification and regression models. The experimental results indicate that the BDA and MBDA have an outstanding ability in knowledge transfer. Compared with support vector machine (SVM) and Gaussian mixture regression (GMR), the proposed methods show a better performance in the decision making of lane keep/change and the prediction of the driver's steering operation.

Fast recognition of a driver's decision-making style when changing lanes plays a pivotal role in a safety-oriented and personalized vehicle control system design. This article presents a time-efficient recognition method by integrating k-means clustering (k-MC) with the K-nearest neighbor (KNN) algorithm, called kMC-KNN. Mathematical morphology is implemented to automatically label the decision-making data into three styles (moderate, vague, and aggressive), while the integration of k-MC and the KNN algorithm helps to improve the recognition speed and accuracy. Our developed mathematical-morphology-based clustering algorithm is then validated by a comparison with agglomerative hierarchical clustering. Experimental results demonstrate that the developed kMC-KNN method, in comparison with the traditional KNN algorithm, can shorten the recognition time by more than 72.67% with a recognition accuracy of 90-98%. In addition, our developed kMCKNN method also outperforms a support vector machine in terms of recognition accuracy and stability. The developed time-efficient recognition approach would have great application potential for in-vehicle embedded solutions with restricted design specifications.

A new driver model adaptation (DMA) method is proposed in this paper to help the model adaptation between different individual drivers. This method is based on transfer learning which can improve the DMA process at data level. The Gaussian mixture model (GMM)-based method is used to model the steering behaviour of drivers during the overtaking manoeuvre. Based on the GMM model, an alignment-based transfer learning technique named local Procrustes analysis (LPA) is modified to formulate the transfer learning problem for driver steering behaviour. A series of experiments based on the data collected from a driving simulator are carried out to evaluate the proposed modified LPA (MLPA). The experimental results verify the ability of MLPA for knowledge transfer. Compared with the GMM-only method and LPA, MLPA shows better performance on the prediction accuracy with much lower predicting errors in most cases.

Establishing a personalized driver model to predict the driving behavior plays a significant role in the promotion of driver assistance system and automated driving system. In this paper, we propose two different learning-based pathtracking personalized driver models to predict the lookahead distance based on pure pursuit algorithm using the naturalistic driving data collected from BIT intelligent vehicle platform. Based on Gaussian Mixture Model (GMM), one stochastic driver model is the velocity-based Gaussian Mixture Regression (GMR) approach established by combining Gaussian classification process and Gaussian mixture regression, while another driver model is the general GMR approach. The predicting results obtained from the stochastic models are analyzed based on the numbers of GMM components. Statistical analyses show that both personalized driver models perform well, and the velocity-based GMR approach demonstrate higher accuracy than general GMR approach in predicting lookahead distance with the preferred number of the GMM components 10-12 and better performance in tracking the given path.

Considering the driving habits which are learned from the naturalistic driving data in the path-tracking system can significantly improve the acceptance of intelligent vehicles. Therefore, the goal of this paper is to generate the prediction results of lateral commands with confidence regions according to the reference based on the learned motion primitives. We present a two-level structure for learning and generalizing motion primitives through demonstrations. The lower-level motion primitives are generated under the path segmentation and clustering layer in the upper-level. The Gaussian Mixture Model (GMM) is utilized to represent the primitives and Gaussian Mixture Regression (GMR) is selected to generalize the motion primitives. We show how the upper-level can help to improve the prediction accuracy and evaluate the influence of different time scales and the number of Gaussian components. The model is trained and validated by using the driving data collected from the Beijing Institute of Technology (BIT) intelligent vehicle platform. Experiment results show that the proposed method can extract the motion primitives from the driving data and predict the future lateral control commands with high accuracy.

This paper develops a learning model for personalized adaptive cruise control that can learn from human demonstration online and mimic a human driver's driving strategies in the dynamic traffic environment. Under the framework of the proposed model, reinforcement learning is used to capture the human-desired driving strategy, and the proportion-integration-differentiation controller is adopted to convert the learning strategy to low-level control commands. The performance of the learning model is tested in the simulation environment built in a driving simulator using PreScan. Experimental results show that the learning model can duplicate human driving strategies with acceptable errors. Moreover, compared with the traditional adaptive cruise control, the proposed model can provide better driving comfort and smoothness in the dynamic situation.