To address the issue of low detection accuracy for small pedestrian targets, a pedestrian detection model tailored for complex scenes is proposed based on the DETR algorithm. In the feature extraction stage, the SENet attention mechanism is introduced to enhance channel discrimination ability. By combining image Relative Position Encoding (iRPE) with absolute position encoding, the model's perception of global context and local details is improved. After training on a self-built dataset, experimental results show that the pedestrian recognition accuracy of the proposed scheme reaches 97.2%, with false detection rate and missed detection rate of only 1.7% and 1.1%, respectively. It performs excellently in handling complex scenes such as multi-scale pedestrians, partial occlusion, and low visibility, and its detection performance is significantly better than traditional methods.

To address the issues of single-type detection and poor adaptability in detecting distant small targets and low-light scenes for road surface irregularity recognition, this paper proposes an improved YOLOv10-based detection algorithm. The SPPF module is replaced with a Focal Modulation network to enhance small-target detection accuracy. The CPA-Enhancer is introduced to improve detection performance under low-light conditions. The Shape-IoU loss function is adopted to optimize localization precision. Experimental results on a self-built dataset demonstrate that the improved model achieves a 3.08% higher mAP@0.5 than the original model. Real-vehicle deployment further verifies the effectiveness of the proposed model, demonstrating its capability for real-time detection of 7 types of road surface irregularities, including speed bumps, potholes, and vibrating markings.

To improve the accuracy and robustness of 3D object detection from a Bird's-Eye View (BEV) in multi-camera visual perception for autonomous driving, this paper proposes a multi-view image object detection method based on an improved BEVFormer. To introduce perspective-based regional auxiliary supervision at the image feature level, using BEV annotations to impose projection constraints on the image view, thereby guiding the backbone network to learn fine-grained local features for 3D spatial localization. This paper establishes instance-level consistency constraints, at the BEV representation level and explicitly enforces the consistency of BEV features for the same physical object across different viewpoints based on a contrastive learning mechanism, thereby mitigating instance-level feature drift caused by view transformations. Furthermore, this paper trains and evaluates the proposed method using the nuScenes dataset and conduct a comparative analysis with various typical multi-view 3D object detection algorithms. The experimental results show that, compared to the BEVFormer baseline model, the proposed method achieves improvements of 1.6 percentage points and 0.4 percentage points in mean Average Precision (mAP) and the Normalized Detection Score (NDS), respectively, while reducing the average localization error by 8.7 percentage points, thereby validating the effectiveness of this method for the BEV detection task using multi-view images.

To address the uncertainty of driving intentions among general vehicles in complex traffic environments, this paper proposes a path planning method based on delayed decision-making. This method generates a trajectory tree comprising an initial shared trajectory segment and branching trajectories corresponding to different potential actions of social vehicles; employs a Gaussian mixture model to predict the probability of social vehicle driving intentions; and determines branching points based on the state deviations of the predicted trajectories. Depending on the requirements of different scenarios, delayed decision-making mechanisms are incorporated into both path planning and speed planning to implement lane-changing and deceleration-based yielding strategies. The results of the joint simulation using MATLAB/Simulink and Carsim indicate that, compared to traditional methods that generate a single trajectory, the proposed approach reduces lateral velocity, acceleration, and their rates of change by 9.2%, 51.7%, and 63.6%, respectively, in lane-changing avoidance scenarios, thereby improving the smoothness of the lane-changing process and ride comfort. In deceleration avoidance scenarios, longitudinal acceleration and its rate of change are reduced by 35% and 61.5%, respectively, significantly improving the smoothness of the deceleration process.

To improve the spatial coverage and timeliness of expressway traffic state estimation, this paper proposes a joint state estimation method that integrates observational data with a traffic flow model based on the Kalman filter. Intelligent Connected Vehicle (ICV) probe data are used as observational inputs. Based on Cooperative Adaptive Cruise Control (CACC) and the Intelligent Driver Model (IDM), the motion behaviors of ICV and human-driven vehicles in mixed traffic flow are simulated, and traffic density is estimated using the Cell Transmission Model (CTM). Validation using SUMO simulations and NGSIM field data shows that the estimation error of the proposed method decreases significantly as the ICV penetration rate increases. Specifically, the Mean Absolute Percentage Error (MAPE) is below 20% at a 20% penetration rate; the estimation accuracy improves further at 40%; and when the penetration rate reaches 80%, the MAPE is reduced by 16.54% compared with statistical estimation methods. The proposed method achieves high estimation accuracy and robustness, which is suitable for traffic state estimation in complex traffic environments.

To address the potential conflicts caused by limited in-wheel space, increased configuration complexity, and the intricate coupling between functions and structures in integrated wheel-side drive-brake systems for electric heavy-duty commercial vehicles, a forward design method for corner-module integrated drive-brake configurations is proposed. First, axiomatic design is employed to analyze the system's functional requirements and structural parameters, while a design matrix is used to identify coupling relationships and conflict issues within the configuration. Then, a systematic design method is applied to construct the morphological matrix of the underlying structures, and potential design schemes are obtained through configuration enumeration. Finally, feasible configurations are selected in accordance with specific application scenarios, and candidate schemes are evaluated using preference matrix analysis to determine the final optional configurations. The results show that the proposed method improves configuration design efficiency from the perspective of functional requirements, assists designers in refining the design process, and provides support for the subsequent establishment of an expert knowledge base in combination with AI-based design.

To address the issue that existing supervised learning methods for estimating the battery State of Health (SOH) rely on scarce labeled data while neglecting the abundance of easily obtainable unlabeled data, a joint SOH estimation method based on Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) within a Semi-Supervised Learning (SSL) framework is proposed. First, three interpretable health indicators are extracted from the charging voltage to accurately describe battery degradation. Then, based on SSL, pseudo-labels for unlabeled data are predicted to augment the training samples, thereby enhancing the generalization capability of the model. Analysis and experiments demonstrate that the proposed SSL-based method significantly improves SOH estimation performance, with an average Root Mean Square Error (RMSE) below 1.2%.

To address the issue that the current single-Bluetooth digital key fails to achieve passive entry due to insufficient positioning accuracy and anti-interference performance, this paper proposes an environment-adaptive dynamic weight allocation algorithm to improve its positioning accuracy and anti-interference capability, thereby realizing stable passive entry in all scenarios. Real-vehicle tests show that with the optimized algorithm, the recognition accuracy reaches 96%, the cost is reduced by approximately 20% compared with most Bluetooth solutions, and it exhibits advantages in both performance and economy.