Instance segmentation algorithms based on deep learning are capable of helping intelligent vehicles to obtain accurate perception information. However, due to the limitation of manufacturing cost, the computing resources on intelligent vehicles are usually limited. In order to obtain highprecision recognition and segmentation under limited computing resources, the algorithm itself is required to make full use of the extracted features. Meanwhile, although the onestage instance segmentation algorithm has a relative fast inference speed, it has poor performance in accuracy. To this end, structural improvement based on the onestage instance segmentation algorithm SparseInst is conducted to enhance the model's utilization of effective features. Specifically, firstly, residual connection is added inside the basic building block of the backbone. Secondly, a threescale feature fusion module is designed to overcome the problem of indirect interaction of crossscale features in the encoder. A decoupled instance activation module is designed to enhance the model's ability to learn instance features. In addition, the improved algorithm makes full use of detail features to refine the mask features to improve the quality of the generated masks. Finally, the kernel is used to initialize the score of the target object, which improves the utilization rate of the extracted features. The improved algorithm surpasses similar algorithms in mask accuracy on multiple datasets and has strong realtime performance. To further verify the effectiveness of the improved algorithm, experiments using data

For the problems of inaccurate speed prediction and poor SOC adaptability under the traditional model predictive control, the plugin hybrid electric vehicle (PHEV) is taken as the research object, and the speed prediction model based on computer vision is combined with the deep deterministic policy gradient (DDPG) algorithm to achieve the realtime state of charge (SOC) reference trajectory planning and optimal power allocation control of PHEV. A SOC reference trajectory planning model based on the enhanced DDPG is constructed, and a speed prediction model based on computer vision with cascaded long shortterm memory network is constructed, based on which the optimal controller based on the model predictive control is used to achieve the accurate tracking of the SOC reference trajectory and power optimization. The results show that compared to the traditional DDPG, the strategy proposed in this paper increases the overall vehicle economy by 5.66%, reaching 97.93% of the global optimal algorithm. It also improves the overall vehicle economy by 2.92% compared to the energy management strategy without computer vision.

Multiclass traffic participant detection in dense traffic scenarios remains a challenging visual task, which is crucial for traffic management and safety. To address this, a deep neural networkbased detection algorithm, DSODet, is proposed to handle the challenges of partial occlusion and smallscale targets in dense traffic environment. Firstly, a lightweight CSPDarkNet network is used to extract features from traffic images. Then, a multiscale feature fusion upsampling module is designed to enhance the representation capability for hardtodetect targets. Next, a highresolution detection branch is incorporated to improve detection accuracy for smallscale targets. Finally, a histogram feature distillation training method is proposed, which effectively guides the student model's training by minimizing the intersection ratio of feature histograms between the teacher and student models at corresponding layers, thus enabling parameter optimization and model compression. The experimental results show that DSODet achieves an average detection accuracy of 66.9% for traffic participants and 13.0% for small targets with partial occlusion, outperforming current stateoftheart algorithms. The model contains only 2.9 M parameters, demonstrating its friendliness for edge device. The related code will be shared at https://github.com/XMUTVsionLab.

Unstructured road often has uneven surface and varying obstacle sizes. Neglecting the uneven terrain and handling obstacles improperly can lead to an imbalance between vehicle safety and travel efficiency. To cope with this challenge, in this paper a trajectory planning method that considers complex terrain and obstacle scales (TOTP) for unstructured road is proposed. Firstly, the trajectoryplanning framework for unstructured road is established based on vehicle passability analysis, to determine the optimal travel pattern. Then, an operational risk field is established based on road roughness and obstacle's size information. In addition, considering both operational risk and travel efficiency, an obstacle avoidance path planning method based on dynamic programming and an obstacle crossing path planning method based on improved A* are proposed. Furthermore, based on vehicle stability analysis, a speed planning method considering terrain constraints is proposed. Finally, realworld experiments are conducted, and the experimental results show that under unstructured road conditions, the trajectory planning method proposed in this paper increases the average vehicle speed by 15.8%, with the average absolute pitch angle and average absolute roll angle reduced by 68.1% and 73.6% respectively. This method can effectively coordinate the safety and efficiency of vehicle operation, demonstrating good generalization and meeting the requirements of realtime performance.

The dualvehicle cooperative transportation system consists of a cargo module and two transport vehicles, which are connected by articulated joints. The system's strong dynamics coupling and nonlinearity present significant challenges for accurate modeling and precise control. In this paper a trajectory tracking control scheme for the cooperative transportation system based on constraintfollowing theory is proposed. In terms of system modeling, based on the kinematics and rigid body dynamics analysis, external trajectory tracking servo constraints and internal articulated passive constraints are constructed for the cooperative transportation system. The Lagrange modeling method is then employed to establish a nonlinear constrained dynamic model of the dualvehicle cooperative transportation system. In terms of controller design, the UdwadiaKalaba (UK) method is first used to obtain the normminimal force required for the cargo to satisfy the trajectory tracking servo constraints, that is, the combined force acting on the cargo at the articulation point. Next, based on the minimum lateral forces principle of front and rear vehicles, an optimal allocation strategy for this combined force is designed, distributing it to the front and rear transport vehicles. The reaction forces of the distributed force components are modeled as the known external disturbances acting on the front and rear vehicles. Then, based on the feedforward compensation for the known external disturbances and the constraintfollowing control theory, the control forces required for the front and rear transport vehicles to satisfy the trajectory tracking servo constraints are designed. Finally, the simulation results show that the proposed cooperative control scheme achieves good trajectory tracking performance and significantly suppresses the lateral dynamic impact of the cargo on the transport vehicles, effectively enhancing the overall lateral stability of the cooperative transportation system.

To improve the accuracy and stability of path tracking for light commercial vehicles under complex curvature conditions, in this paper a PredictivePure Pursuit (PPP) control method is proposed. Firstly, a PPP controller is designed based on the vehicle's discrete kinematic model, and a PID compensator is developed based on heading error to enhance tracking accuracy and stability. Secondly, to address the challenge of maintaining both accuracy and stability under complex curvature conditions with a fixed prediction horizon algorithm, a variable prediction horizon optimization algorithm is proposed. A cost function based on the lateral and curvature errors within the prediction horizon is established, and Bayesian optimization is used to determine the optimal prediction horizon, resolving the conflict between accuracy and stability. Finally, TruckSim/Simulink cosimulation and real vehicle tests are conducted. In the real vehicle tests, the root mean square values of the lateral error, heading error, and steering wheel angle for the Bayesianoptimized PPP controller is 0.113 m, 0.045 rad, and 153.2°, respectively, all of which are superior to the corresponding metrics of the PPP controller based on fuzzy control and the MPC controller, indicating that the proposed controller maintains good precision and stability under complex curvature conditions.

Threedimensional terrain scenes typically possess complex and diverse environment along with jagged terrain features, which poses challenges to path planning. To address this issue, in this paper a highly reliable path planning approach under the influence of nontopographic fluid characteristics and threedimensional complex terrain is proposed. This method encompasses initial global path planning, path inspection, and replanning under 3D terrain. For the initial global path, an AHTR algorithm that combines the advantages of Hybrid A* and Theta* is proposed. This algorithm enhances the internode sampling and detection methods in accordance with the traits of the 3D terrain scene and introduces in a terrain risk assessment function.to plan a path for the vehicle that can evade rough terrain and comply with kinematic constraints. For path inspection, the path risk test function is designed based on the results of vehicle dynamics analysis considering the characteristics of nonterrestrial fluid, and the impact of nonterrestrial fluid characteristics on path planning is verified. For path replanning, an enhanced AHTR algorithm is proposed, which takes into account of both 3D terrain features and nonterrain fluid features to guarantee that the planned path can effectively avoid risks. Simulation experiments demonstrate that compared with Hybrid A* and Theta*, the intensity of ground undulation in the path planned by the AHTR algorithm is decreased by 26.54% and 49.04%, with the average pitch angle of the vehicle reduced by 44.39% and 69.40%, the path risk lowered by 26.32% and 41.67%, and the final path safety improved by 58.06% and 88.46%, which effectively ensures path reliability.

Injection overmolding after compression for thermoplastic composites balances the low cost and high performance. It can quickly and stably achieve integration of continuous/discontinuous fiberreinforced composites, meeting the requirements of the automotive industry. This paper reviews the key points of equipment selection, process control and molding simulation, introduces the types of domestic materials, summarizes the optimization design methods for anisotropic bimaterial structures, generalizes the evaluation methods for interface and overall performance, and puts forward some critical scientific and technological issues in the integrated design of "material process structure performance".

For the high energy consumption problem caused by the large torque fluctuations of inwheel motordriven vehicles on uneven roads and frequent shifting of vehicles, in this paper a method for energy consumption optimization of inwheel motordriven vehicles considering torque fluctuations is proposed. Firstly, based on the longitudinal drive dynamics equation of electric vehicles and the CarSim vehicle dynamics model, the motor energy consumption model, tire slip energy consumption model, and yaw torque tracking error model are established as the upper control target, and the inwheel motor dynamics model considering road surface excitation is established as the lower control target. Secondly, taking the upperlevel control target as the objective function of the torque optimization of the inwheel motor vehicle, the motor torque and speed energy limit as the inequality constraints, and the fuzzy control method for objective function weight distribution, the upperlevel torque optimization model is established based on NSGA II. At the same time, the lowerlevel inwheel motor vector control model of torque overshoot and delay caused by road surface excitation is established based on the sliding mode antidisturbance observer. Finally, the joint simulation of Simulink and CarSim of a fourwheel inwheel motordriven car is carried out, and the changes of the front and rear axle wheel torque, total energy consumption of the car, and the SOC of the car battery under different optimization methods are analyzed under WLTC operating conditions and CLTCP operating conditions. The inwheel motor bench test shows that the energy consumption optimization method of inwheel motordriven vehicles considering torque fluctuations can effectively reduce energy consumption under WLTC and CLTC-P operating conditions.

In the realm of vehicle dynamics, the sideslip angle is a critical parameter. For the challenges posed by the current modelbased methods, which heavily rely on the accuracy of dynamic models, and the poor robustness of datadriven methods in unfamiliar operating conditions, in this paper a sideslip angle estimation method based on a hybrid of physics and datadriven approaches (DeepPhy) is proposed. The aim is to combine the strength of physical modeling and datadriven techniques to achieve reliable and accurate estimation of the sideslip angle. DeepPhy integrates prior values of the sideslip angle obtained from the lateral force model of the rear axle tires with a deep neural network, enabling the learning of nonlinear mapping relationship not captured by the physical model, thereby enhancing the model's reliability in unfamiliar conditions. The simulation results indicate that under continuous DLC conditions, the RMSE of the estimation results from DeepPhy is reduced by 93% compared to the physical model method and by 63% compared to the datadriven method, exhibiting robustness in scenarios with limited data. Realworld validation further confirms DeepPhy's exceptional generalization capabilities, as the models trained through simulation can be transferred to realworld conditions while maintaining highprecision estimation results.