To address the issues of low efficiency, poor quality of random sampling points, and suboptimal path quality in the traditional RRT* algorithm during path exploration, this paper proposes a method that integrates an improved artificial potential field formulation with an adaptive step-size bidirectional expansion mechanism. By introducing an adaptive attraction coefficient and dual repulsion forces, the proposed method optimizes path guidance and obstacle avoidance. A dynamic step-size adjustment strategy and a bidirectional tree cooperative search mechanism are employed to reduce redundant nodes and accelerate convergence. A greedy algorithm is used to remove redundant nodes, ensuring path smoothness and continuity. Additionally, an adaptive goal-biasing strategy based on environmental density is introduced to enhance goal orientation and exploration efficiency. Experimental results demonstrate that, compared with the traditional RRT* algorithm and other improved algorithms, the proposed ABF-RRT* algorithm reduces the average path length by 17.7%, decreases the average number of nodes by 90.6%, and shortens the average running time by 95.5% in complex environments, verifying its stability and superiority.

To address the issues of insufficient accuracy and limited speed of the RT-DETR (real-time detection transformer) network in complex scenarios and small object detection, an improved detection algorithm for vehicle images is proposed. This algorithm enhances the ability to capture features of small-scale objects such as vehicles and pedestrians by introducing spatial positional relationship modeling and improves the network's nonlinear modeling performance through a gating mechanism. At the same time, by separating high-frequency and low-frequency detail information and combining feature map segmentation and stitching operations, feature sampling is achieved without information loss. By integrating convolution units (Conv) and batch normalization units (BN) to reduce intermediate processing steps and memory consumption, computational efficiency is increased and model performance is optimized. Experimental results show that compared to the original RT-DETR network, the improved algorithm increases mAP50, mAP50∶95, precision, and recall by 6%, 4.5%, 4.2%, and 4%, respectively. Therefore, optimizing the backbone network structure and strengthening multi-scale feature fusion, focusing on fine-grained information and spatial-frequency domain features, is an effective strategy to improve vehicle detection performance in complex environments.

To explore the traffic violation characteristics of Autonomous Vehicles (AV) in real-world traffic environments, the adaptability of existing traffic regulations, and the differences compared to Human-Driven Vehicles (HV), this study utilizes the SinD artificial driving dataset, highD artificial driving dataset, and Argoverse2 autonomous driving dataset. 6 types of violation cases are analyzed, including speed limit violations, safe following distance violations, lane changes, running red lights, left-turn blocking of straight-moving vehicles, and continuous lane changes. The study employs Pearson correlation coefficients and linear regression models to analyze speed limit violations, evaluates the safety of following distance and lane changes using headway time as an indicator, and uses data proportion analysis to assess red light violations and left-turn blocking violations that impede straight-moving traffic. Additionally, continuous lane change characteristics are studied by examining lane change frequency, stability, and speed variability. The results show that under high-speed conditions, AVs are more prone to persistent speed limit violations but exhibit more stable speeds. The risk of safe following distance violations is higher for AVs compared to HVs, while the risk of lane change violations is lower for AVs. Red light violations are more significantly influenced by weather conditions for AVs, and once left-turn blocking violations occur, they present more severe risks. In short-distance, low-speed scenarios, AVs tend to change lanes more frequently with poorer stability. The study confirms that AVs have not yet fully complied with traffic regulations, with higher rates of speed limit and safe following distance violations. Significant differences exist between AVs and HVs across all six types of violations. The applicability of current traffic regulations to AVs has limitations. Based on this, future research can further explore 4 areas: the construction of specialized AV violation datasets, the improvement of specific traffic regulations for AVs, upgrades to autonomous driving algorithms, and the establishment of compliance evaluation mechanisms. The findings provide theoretical support and practical reference for the compliant integration of AVs into the existing traffic system.

To address the impact of blurred environmental features caused by snow and ice coverage on LiDAR point clouds during real-time laser-based construction and Simultaneous Localization and Mapping (SLAM) in extremely cold environments, this paper employs the LIO-SAM algorithm framework. In the point cloud preprocessing stage, a Dynamic Radius Outlier Removal (DROR) filter is used to eliminate snow-induced noise while preserving the details of environmental features, thereby providing high-quality point cloud data for subsequent processing by the LIO-SAM algorithm. Validation results on the CADC dataset show that, compared to the filter built into LIO-SAM, the point cloud data collected under snowy conditions achieves a recall rate exceeding 91% while maintaining high accuracy. This approach effectively removes snow-induced noise while preserving more environmental features.

To address the critical issues of range anxiety among battery electric vehicle users and the insufficient accuracy of energy consumption prediction, this paper proposes a two-stage energy consumption prediction method considering driving styles. First, characteristic parameters strongly related to driving styles, such as speed, accelerator pedal depth, and brake pedal depth, are extracted and calculated, and a back propagation neural network is used to construct a driving style recognition model. Second, for different driving styles, decision tree models are respectively used for energy consumption prediction. Verified by real vehicle test data, the average absolute percentage error (MAPE) of the energy consumption prediction model proposed in this paper is 5.84%. Compared with the average energy consumption method, the Deep Neural Network (DNN) and decision tree energy consumption prediction methods that do not consider driving styles, the errors are reduced by 16.74%, 9.77% and 1.44% respectively. The research results demonstrate that the energy consumption prediction method incorporating user driving style can effectively alleviate user range anxiety and significantly improve the accuracy of energy consumption prediction.

To address the insufficient accuracy of State Of Health (SOH) prediction for power batteries under real-vehicle operating conditions, a prediction method integrating Temporal Convolutional Network (TCN) and Transformer is proposed. Using real-vehicle operation monitoring data as the research object, after removing abnormal values and selecting charging condition data, the current battery capacity is calculated via the ampere-hour integration method. The Whale Optimization Algorithm (WOA) is adopted to optimize the parameters of Variational Mode Decomposition (VMD), realizing the decomposition and denoising of the capacity sequence. Features are extracted by combining parameters such as current and temperature, and 6 strongly correlated features are selected to construct a dataset, which is input into the model to complete SOH prediction. Results show that the model significantly reduces prediction errors in the datasets of 4 real-vehicle batteries, with the coefficient of determination [R²>0.8], providing technical support for the full-life-cycle health management of power batteries.

To address the complex operating conditions and diverse fault types of electric mining truck drive motors, a fault diagnosis method integrating the Caterpillar Fungus Optimizer (CFO) with Support Vector Machine (SVM) is proposed. A parameter optimization model is constructed to jointly optimize the slack variable, penalty factor, and kernel function parameter of SVM, thereby improving fault recognition accuracy and model stability. Fault sample data are collected under typical operating conditions for identification experiments, and uniformly distributed random disturbances with different amplitudes are introduced to evaluate model robustness. Comparative analysis with several commonly used optimization methods demonstrates that the proposed approach exhibits strong adaptability in terms of fault classification accuracy, anti-interference capability, and model training stability. The average recognition accuracy reaches 98.67%, while the accuracy remains above 94.7% under disturbance conditions.