Rule-based strategies often fail to respond quickly under complex conditions, leading to increased energy consumption. Advanced energy management strategies (EMSs) serve as critical determinants of operational safety-performance metrics (durability, reliability, and structural integrity) in new energy vehicles. To address these requirements, this study develops a velocity prediction (VP) EMS incorporating deep reinforcement learning method, specifically designed to resolve autonomous mode switching issue in a novel electro-hydraulic coupled hybrid vehicle (EHCHV). A driving intention is integrated into the vehicle speed prediction method. Then VP-DDQN-EMS is proposed by combining the velocity prediction with the double depth Q-network (DDQN) EMS. Finally, a test bench is built to validate the reliability of EMS and vehicle model. Compared with electric vehicles and rule-based EMS (RB-EMS), under unchanged high-pressure accumulator (HPA) recovery efficiency, the VP-DDQN-EMS achieves 4.43%, 5.21%, and 7.47% higher energy-saving rates than RB-EMS in NEDC, CLTC-C, and real cycle conditions respectively. The VP-DDQN-EMS achieves lower mode switching frequency, stable transition and less transient mutation in a very short time. Through multi-scenario power source coordination mechanism, combined with real-time road conditions, it dynamically optimizes the balance between driving performance, power distribution and energy efficiency. This approach achieves a breakthrough in overcoming the mode-switching accuracy limitations of traditional electro-hydraulic hybrid vehicles in complex scenarios.

Abstract Recent advances in vision–language models (VLMs) have opened new opportunities for interpretable decision-making in autonomous driving, yet directly deploying VLMs for real-time control remains impractical due to high computational cost and the semantic gap between abstract language and continuous actions. To address these challenges, we present DualDrive , a dual-system driving framework inspired by dual-process theory. System 2 employs a VLM as a low-frequency reasoning module that interprets human or map-based instructions and outputs probabilistic trajectory distributions. System 1 functions as a high-frequency controller that consumes visual observations, vehicle states, and the embedding of a dedicated token. This token is explicitly designed for autonomous driving, serving as a compact semantic channel that encodes driving intentions from high-level reasoning and delivers them to the low-level controller for precise execution. This principled separation achieves a balanced trade-off between interpretability and real-time responsiveness. To evaluate our approach, we conduct extensive experiments on the Bench2Drive benchmark, covering 220 routes across diverse traffic, weather, and safety-critical scenarios. DualDrive achieves state-of-the-art performance in both open-loop and closed-loop evaluations, significantly reducing trajectory prediction error while attaining higher Driving Score and Success Rate compared to strong baselines. Furthermore, case studies demonstrate that the proposed framework enables interpretable, human-aligned decision-making that adapts effectively to rare and safety-critical events.

In the domain of autonomous driving systems, vehicle trajectory prediction represents a critical aspect, as it significantly contributes to the safe maneuvering of vehicles within intricate traffic environments. Nevertheless, a preponderance of extant research efforts have been chiefly centered on the spatio-temporal relationships intrinsic to the vehicle itself, thereby exhibiting deficiencies in the dynamic perception of and interaction capabilities with adjacent vehicles. In light of this limitation, we propose a vehicle trajectory prediction algorithm predicated on a hybrid prediction model. Initially, the algorithm extracts pertinent context information pertaining to the target vehicle and its neighboring vehicles through the application of a two-layer long short-term memory network. Subsequently, a fusion module is deployed to assimilate the characteristics of the temporal influence, spatial influence, and interactive influence of the surrounding vehicles, followed by the integration of these attributes. Ultimately, the prediction module is engaged to yield the predicted movement positions of the vehicles, expressed in coordinate form. The proposed algorithm was trained and validated using the publicly accessible datasets I-80 and US-101. The experimental results demonstrate that our proposed algorithm is capable of generating more precise prediction results.

For lane change behavior under extreme operating conditions, existing models cannot calculate in real time the tire force of the vehicle lane change over a sufficiently long time frame in the future. In order to address this problem, a novel scheme is presented for real-time trajectory planning of autonomous vehicles, which incorporates personalized vehicle dynamics. We first establish lateral dynamics models for four-wheel-steering and front-wheel-steering vehicles along with a nonlinear tire model. Then, we construct a fuzzy logic mechanism to characterize the relationship between the vehicle lateral/longitudinal acceleration and the future lateral/longitudinal tire force, to quantify whether the vehicle tire force reaches saturation in trajectory planning in real time. A safety assessment model is introduced to measure the risk of side slippage of the vehicle and collision under extreme operating conditions. In addition, lane change behavior is designed as a nonlinear programming model and a gradient descent method is used to obtain optimal lateral and longitudinal accelerations online. The geometric curve fitting method is utilized to generate the lane change trajectory. The simulation results using MATLAB/Simulink demonstrate that the solution time of our method is significantly lower than that of the widely used vehicle dynamics method and the newest Neural Network method, which can realize real-time prediction of the maximum tire force before lane change. Moreover, our method improves the ability to calculate the risk of longitudinal and lateral coupling of a lane change in extreme operating conditions and then realizes trajectory planning in a vehicle-dynamics-specific way.

Emergency collision-avoidance maneuvers demand rapid and coordinated vehicle responses, especially when both longitudinal and lateral dynamics become highly coupled. Conventional trajectory tracking controllers-typically based on simplified single-track or constant-speed models-often neglect these coupling effects, resulting in degraded control accuracy and stability under high-speed evasive conditions. To address these challenges, this paper proposes a linear time-varying model predictive control (LTV-MPC) framework for emergency collision avoidance and trajectory tracking of four-wheel-steering (4WS) vehicles. The controller is derived from a nonlinear seven-degree-of-freedom vehicle dynamics model and linearized around the current operating point to ensure real-time solvability. It jointly regulates longitudinal driving force and steering angle to minimize tracking errors along a preplanned evasive trajectory. A closed-loop simulation platform based on Carsim and MATLAB/Simulink is established to validate the control performance under double lane-change and sudden obstacleavoidance scenarios. Results demonstrate that the proposed LTVMPC achieves superior trajectory tracking accuracy, stability, and computational efficiency compared with conventional DPID-LQR and DPID-MPC controllers, providing a feasible real-time solution for safety-critical vehicle maneuvers.

The widespread use of pesticides poses many potential risks to food safety and human health. Thus, rapid and accurate detection methods for pesticide residues need to be established. In this study, ultraviolet (UV) spectroscopy coupled with support vector regression and variable selection methods was used to quantitatively detect the content of imidacloprid in apple juice. First, the UV spectra of diff erent imidacloprid concentrations in apple juice were collected, and the acquired spectra were preprocessed by Savitzky-Golay smoothing. Then, the feature variables were selected by the variable iterative space shrinkage approach (VISSA), iteratively retains informative variables (IRIV), and random frog (RF) algorithms. Finally, particle swarm optimization support vector regression (PSOSVR) prediction models based on the feature variables and the full-spectrum variables were established to detect imidacloprid in apple juice. The results showed that the VISSA-PSO-SVR model had the optimal predictive performance, the determination coefficient of the prediction set (Rp2) was 0.99933, and the root mean square error of the prediction set (RMSEP) was 0.0894 mg/L. The results from this study indicated that the combination of UV spectroscopy and the VISSA-PSO-SVR model could be used for the quantitative detection of imidacloprid in apple juice.

The distribution of tags is an important factor that affects the performance of radio-frequency identification (RFID). To study RFID performance, it is necessary to obtain RFID tags' coordinates. However, the positioning method of RFID technology has large errors, and is easily affected by the environment. Therefore, a new method using optical measurement is proposed to achieve RFID performance analysis. First, due to the possibility of blurring during image acquisition, the paper derives a new image prior to removing blurring. A nonlocal means-based method for image deconvolution is proposed. Experimental results show that the PSNR and SSIM indicators of our algorithm are better than those of a learning deep convolutional neural network and fast total variation. Second, an RFID dynamic testing system based on photoelectric sensing technology is designed. The reading distance of RFID and the threedimensional coordinates of the tags are obtained. Finally, deep learning is used to model the RFID reading distance and tag distribution. The error is 3.02%, which is better than other algorithms such as a particle-swarm optimization back-propagation neural network, an extreme learning machine, and a deep neural network. The paper proposes the use of optical methods to measure and collect RFID data, and to analyze and predict RFID performance. This provides a new method for testing RFID performance.

On highways, the interaction with surrounding vehicles is very crucial for the decision-making and planning of autonomous vehicles. However, the multi-modal driving intentions of surrounding vehicles have brought great challenges. Aiming at the multi-modal driving intention of surrounding vehicles, a multi-modal driving risk field based on dynamic collision region is proposed, and multi-modal driving intention partially observable markov decision process (MDI-POMDP) decision framework is established, which integrating behavior decision and motion planning. Firstly, the multi-modal probability distribution of driving intention is fused to establish a driving risk field. Moreover, combined with the longitudinal safety distance model and lateral driving direction, the concept of dynamic collision area is proposed in the driving risk field. Then, MDI-POMDP is formulated to analyze the influence of the uncertainty on planning, which is caused by the multi-modal driving intention of surrounding vehicles. In the following, with the help of the previous state, a time-dependent deep reinforcement learning (DRL) algorithm recurrent deterministic policy gradient (RDPG) is designed to enhance the current observation, to solve the optimal driving policy under partial observation and generate the optimal trajectory. Furthermore, the simulation results show that the performance of our proposed motion planning algorithm is outstanding, compared with the states-of-the-art methods. And our algorithm has the powerful ability to model the multi-modality of driving intention, to ensure the traffic safety.

The on-ramp area usually produces congestion, high energy consumption, and high emission. In order to improve the efficiency of multi-lane heterogeneous on-ramp traffic composed of vehicles with different dynamic characteristics, we propose an eco-friendly on-ramp merging strategy for connected and automated vehicles in heterogeneous traffic. Firstly, the optimal lane decision method is proposed via the conditional proximal policy optimization algorithm to optimize the traffic flow of each lane and avoid local congestion in heterogeneous traffic. Then, considering travel time, energy consumption, and emission, the eco-friendly merging optimization problem is established to optimize vehicles’ longitudinal velocity. According to the optimal lane and the optimal longitudinal velocity profile, the lateral end-point and longitudinal end-point of the local trajectory are solved for planning the local trajectory of the vehicles. Each vehicle tracks its local trajectory, realizing eco-friendly on-ramp merging. Compared with two existing ramp merging methods, results show that the proposed algorithm can mitigate local congestion effectively, and has superiority in traffic efficiency, energy economy, and eco-friendliness.

摘要: 为了解决无信号交叉路口自动驾驶车辆决策保守,与周围车辆交互性差的问题,提出一种考虑交互博弈的无信号交叉路口自动驾驶车辆决策规划算法。该方法分为以下几步,首先基于运动学模型及道路约束对自车和周围车辆进行初步的运动预测,并建立两车的交互动作空间,得到车辆可能的行驶域。其次,建立一个新颖的危险度评估方法,能评估两车在任意位置、任意姿态、任意速度下的碰撞危险度,用于各交互动作状态行为值的求解。进一步基于斯塔克伯格主从博弈求出两车的均衡动作策略,该策略即为当前交通环境下考虑交互得到的最优动作。最后,通过Prescan/Simulink构建交叉路口场景进行联合仿真,来验证该算法的合理性。结果表明所提出的考虑交互博弈的算法在保证安全性的基础上,能相对于基于决策树和无交互的方法分别提高7.3%和12.4%的效率,并能在多车复杂工况下与周围车辆进行灵活交互。

Uncertain cut-in maneuver of vehicles from adjacent lanes makes it difficult for vehicle's automatic speed control strategy to make judgments and effective control decisions. In this paper, an intelligent speed control strategy for uncertain cut-in scenarios is established based on a basic autonomous driving system. This strategy judges cut-in maneuver from surrounding vehicles and outputs adaptive control action under current environment according to Q value of state-action pair based on a Q network. In addition, according to the analysis of cut-in scenarios, the Q network is trained based on a novel reinforcement learning method named as experience screening deep Q-learning network (ES-DQN). The proposed ES-DQN is an extension of double deep Q-learning network (DDQN) algorithm, and includes two parts: experience screening and policy learning. Based on the experience screened from the experience screening part, the proposed learning method can train an intelligent speed control strategy which has stronger adaptability and control effect in uncertain cut-in scenarios. According to simulation results, the proposed intelligent speed control strategy trained by ES-DQN has better performance under uncertain cut-in scenarios than DDQN method and traditional ACC strategy. Meanwhile, by adjusting weight value in reward function, the system can realize different control target.

Trajectory prediction plays a key role in the decision-making system of autonomous vehicles. The existing trajectory prediction models have the problem of accuracy deterioration when driving behavior change. To solve this problem, a novel BRAM-ED trajectory prediction framework is proposed, which mainly consists of driving behavior recognition model, behavior attention mechanism (BAM) trajectory encoder, and behavior adaptive future trajectory decoder. First, a Bi-GLSTM network is designed to recognize real-time driving behavior. The graph structure is used to describe the complex dependencies among vehicles and recurrent cell is to capture temporal correlation. Subsequently, in contrast to traditional models, directly concatenate historical trajectories and interactive information as the input of prediction model, BAM is proposed to integrated trajectory and interaction features in this article. The BAM is designed to generate attention weights according to the change of driving behavior, and guide trajectory decoder to predict the future trajectories. Then, the proposed model is trained and validated on various public datasets, including HighD and NGSIM. Compared with existing optimal models, our prediction error has been reduced 44.66% at most. Furthermore, a typical cut-in scenario is designed based on hardware-in-loop platform. The experiment results show that the proposed model could recognize the change of driving behavior timely, and predict accurate trajectory.

为实现智能车对各种复杂路径快速稳定的识别,设计一种根据视觉图像进行路径识别的智能车控制系统.该智能车控制系统以MK60DN512ZVQ10控制器为核心,采用OV7725视觉传感器获取赛道的二值化图像;通过图像处理提取赛道边沿及中心线,并提出环岛等复杂路径的识别算法;通过增量式编码器测量实时车速,采用PID控制算法控制舵机的转向和驱动电机的转速;通过调整视觉传感器的高度并合理布局车体结构,增强智能车的路径识别能力和稳定性.实验测试结果表明:利用该系统,智能车在速度约为2.4 m/s时可在赛道上快速稳定行驶并正确识别各种路径.

Recently, inferring lane change intention has received considerable attention. Due to the high nonlinearity and complexity of traffic contexts, traditional methods cannot satisfy the requirements of long-term prediction tasks and lack the ability of capturing nonlinear temporal dependencies. This paper proposes an intention inference model based on Recurrent Neural Networks (RNN), to tackle time series prediction problems. Considering dynamic interaction among surrounding vehicles, our model takes the sequence motion information of surrounding vehicles as inputs and calculates the congestion of different lanes, integrated with vehicle states of the object vehicle. To illustrate the availability of the proposed RNN intention inference model, a motion planning controller considering intention was developed. A Nonlinear Model Predictive Control (NMPC) was established to planning a safe, sub-optimal path for autonomous driving vehicle under collision avoidance constraints. The experiments on the proposed model were conducted, based on two RNN structure Long-Short Term Memory (LSTM) and Generalized Recurrent Unit (GRU), by Tensorflow with NGSIM data. The motion planning controller is modeled and simulated by Carsim with Simulink for some typical scenarios. Subsequently, experimental results demonstrate that RNN achieves best performance, inferring intention with 96% accuracy, compared with other approaches.

Multitag sensitivity would be affected by electromagnetic coupling during simultaneous reading. The reading distance of multitag depends on the least sensitivity and operating power of all tags, which is an indicator of reading performance. The main purpose of this article is to optimize the reading performance of multitag by embedding deep learning, while the reading distance of multitag changes with the 3-D geometric structure. Dynamic multitag image deblurring based on multiscale convolutional neural network (MCNN) improves image restoration ability and clarity. Also, tag detection from the estimated image via YOLOv3 improved by feature enhancement (YOLOv3_f) can improve the detection ability of small size targets, and mean average precision (mAP) is increased by 16.4%. Finally, the 3-D coordinates of tags in pixel space are converted into 3-D coordinates of world space by a quaternion. Comparing our system with the positioning method without deblurring, the 3-D coordinate structures are tested in the dynamic measurement system. The experimental results show that the reading performance of the designed RFID system has been greatly improved as the number of tags increases. Our scheme can improve the reading distance of multitag from the physical structure and the anticollision ability of the RFID system.

The prediction and estimation of the lane-changing state of the host car and surrounding cars are important parts of an advanced driving assistant system, which mainly depend on the understanding of the driver lane-changing behavior. To learn driver lane-changing maneuver well, this article provides a novel stochastic driver lane-changing model based on an improved input-output hidden Markov model (IOHMM) framework. First, an improved IOHMM is proposed to address the deficiency that the traditional IOHMM cannot remember previous data and describe continuous output. Then, based on the improved IOHMM framework, a driver lane-changing model is established considering the intention and behavior of the driver in the lane-changing process. The model parameters can be learned from the collected lane-changing data using the maximum likelihood estimation and generalized estimation-maximization methods. Finally, the model is applied to a real driver lane-changing process. It is verified that the proposed model has good performance in predicting the future motion maneuver of the host vehicle and estimating the current motion state of the surrounding cars.

Under the actual operating conditions of electric vehicle, the temperature of the battery system changes with the ambient temperature, its influence is rarely considered on state of charge (SOC) estimation. Aiming at this problem, this work explores the relationship between temperature and battery discharge capacity, capacity loss, coulomb efficiency and open circuit voltage, and establishes a temperature correction model. Then, a temperature-based coulomb counting (TBCC) method is proposed. To eliminate the cumulative errors of SOC estimation, a TBCC combined adaptive particle filter (APF) combination estimation method is proposed, which can update the temperature-related variables, and estimates the system state with the identified model parameters. The estimation method adapts to the temperature change of the power lithium battery while taking into account the superiority of the APF in non-linear and non-Gaussian systems, so as to improve the accuracy of the SOC estimation in the changing temperature range. Simulations under urban dynamometer driving schedule (UDDS) are conducted, and the results show that compared with the method without temperature correction, the proposed combination method can improve the accuracy and convergence ability of SOC estimation under the influence of temperature.

In order to improve the efficiency and comfort of autonomous vehicles while ensuring safety, the decision algorithm needs to interact with human drivers, infer the most probable behavior and then makes advantageous decision. This paper proposes a Nash-Q learning based motion decision algorithm to consider the interaction. First, the local trajectory of surrounding vehicle is predicted by kinematic constraints, which can reflect the short-term motion trend. Then, the future action space is built based the predicted local trajectory that consists of five basis actions. With that, the Nash-Q learning process can be implemented by the game between these basis actions. By elimination of strictly dominated actions and the Lemke-Howson method, the autonomous vehicle can decide the optimal action and infer the behavior of surrounding vehicle. Finally, the lane merging scenario is built to test the performance contrast to the existing methods. The driver in loop experiment is further designed to verify the interaction performance in multi-vehicle traffic. The results show that the Nash-Q learning based algorithm can improve the efficiency and comfort by 15.75% and 20.71% to the Stackelberg game and the no-interaction method respectively while the safety is ensured. It can also make real-time interaction with human drivers in multi-vehicle traffic.