Monocular 3D object detection is critical for roadside perception but faces significant challenges due to high camera mounting positions and large pitch angles, which cause strong perspective distortion and scale-resolution variation, especially for distant objects. Under fixed roadside cameras on mostly planar roads, image coordinates offer approximate perspective cues for depth-related modeling, but are not universally valid geometric constraints. To address this, we propose MonoCAP, a Coordinate-Aware and Perspective-Aware Learning framework. Specifically, we design a Coordinate-Aware Hybrid Vision Encoder (CAHVE) that incorporates coordinate-aware positional cues through 2D positional embeddings and Coordinate Attention, enhancing spatially sensitive feature representation for distant objects. Furthermore, a Geo-Visual Synergistic Attention (GVSA) mechanism is introduced to learn a data-driven fusion ratio between ground-aware geometric features and visual semantic features, providing a more flexible geo-visual aggregation strategy than manually fixed fusion. Additionally, we construct a Perspective-Aware Deformable Attention (PADA) module, which introduces vertical-coordinate cues as soft perspective guidance for multi-scale feature selection and spatial sampling, improving feature adaptation to the near-large-far-small scale variation under fixed roadside viewpoints. Extensive experiments on the DAIR-V2X-I and Rope3D benchmarks demonstrate that MonoCAP achieves competitive performance and particularly improves the detection of distant objects under typical roadside perspective variations.

In autonomous driving decision-making systems, conventional single-mode approaches—either knowledge-driven or data-driven—each have inherent strengths and limitations. Knowledge-driven methods provide high interpretability but lack adaptability in complex scenarios, whereas data-driven methods exhibit strong learning capability while relying heavily on large amounts of labeled data and offering limited interpretability. At present, hybrid decision-making frameworks that integrate cognitive modeling and data learning are widely regarded as an effective means to address these limitations. However, the deep integration of knowledge and data still faces challenges in theoretical foundations and practical implementation. To this end, this paper proposes a hybrid behavior decision-making framework based on the adaptive control of thought—rational (ACT-R) cognitive architecture and deep reinforcement learning. First, an optimized decision tree method is used to automatically construct the procedural knowledge module within ACT-R, replacing traditional manual rule construction to improve modeling efficiency and generalization capability. Second, a hybrid decision-making mechanism that combines ACT-R cognitive strategies with a proximal policy optimization (PPO) policy is designed to enhance autonomous learning capabilities while ensuring safety during training. Third, an adaptive clipping strategy is introduced to dynamically adjust the PPO clipping factor according to the source of each strategy, thereby improving training stability and policy performance. Experimental results show that the proposed ACTR–ADPPO method outperforms existing comparison algorithms in terms of convergence speed, reward values, and safety performance, demonstrating the effectiveness and superiority of the hybrid decision-making framework in complex autonomous driving scenarios.

Accurate motion prediction of traffic participants is essential for the safe planning of autonomous vehicles. However, conventional interaction modeling often lacks a hierarchical information transfer mechanism, making it difficult to capture deep dependencies between agents and road topology. Furthermore, traditional direct regression or static anchor-based decoding strategies struggle to adapt to diverse driving intentions and lack mechanisms for geometric refinement based on scene context. To address these issues, this paper proposes a trajectory prediction model based on hierarchical spatial-aware interaction and intent guidance. First, a hybrid feature encoder is constructed by integrating spatio-temporal attention with State Space Models (SSMs) to efficiently extract long- and short-range contextual features from dynamic scenes. Second, a hierarchical bidirectional spatial interaction mechanism is developed. By establishing multi-round information flows between agents and lane lines, the model achieves hierarchical feature fusion ranging from global scene perception to local fine-grained interaction. Finally, an intent-guided two-stage decoding strategy is designed: the first stage generates initial anchors incorporating intent priors via a multimodal network, while the second stage applies geometric bias correction based on scene context to produce refined trajectories that adhere to physical constraints. Experimental results on the Argoverse 1 and 2 benchmarks demonstrate that the proposed model achieves a superior balance between prediction accuracy and inference speed, validating the effectiveness of the hierarchical interaction and two-stage decoding strategies.

Remanufacturing scheduling is a critical path to sustainability. However, existing research on the three-stage remanufacturing system scheduling problem (3T-RSSP) is predominantly concentrated in the deterministic domain; the few studies involving uncertainty mainly employ stochastic programming or symmetric fuzzy sets, neglecting the asymmetric uncertainty inherent in remanufacturing. This paper investigates the bi-objective energy-aware 3T-RSSP under asymmetric uncertainty, integrating disassembly, reprocessing, and reassembly. First, a fuzzy mathematical model is formulated to minimize makespan and energy consumption; by quantifying aggregate asymmetric uncertainty via a Right-Skewed Perturbation Model based on Triangular Fuzzy Numbers, it captures operational delay characteristics to accurately map real-world scenarios. Second, an Uncertainty-Aware Scheduling Strategy (UASS) is proposed to preserve the stability of the scheduling scheme. By treating tasks with similar expected processing times as indistinguishable ones and using the Coefficient of Variation as a tie-breaking criterion, UASS implements a risk postponement mechanism to shift high-volatility operations to later stages. Third, an Improved Hybrid Guided Whale Migration Algorithm (HGWMA) is proposed, incorporating AGALF and PAVM strategies to balance exploration and exploitation. Finally, comprehensive simulations across varying scales evaluate HGWMA against other algorithms. Compared with NSGA-II, HGWMA achieves a 12.8% average Hypervolume improvement and significantly reduces Inverted Generational Distance, especially in large-scale instances. Results demonstrate the feasibility and superiority of the proposed framework in addressing the energy-aware scheduling problem under asymmetric uncertainty.

Few-shot object detection aims to train a model using only a limited number of labeled samples. In such scenarios, the model is expected to rapidly learn the distinguishing features of new categories from the scarce data. Previous meta-learning-based approaches typically extract class-level prototypes from support samples and then fuse them with query features to obtain the final aggregated features. However, these prototypes are commonly obtained by averaging multiple labeled samples from the same class, which treats the samples as equally important. Therefore, it is easy to learn redundant features and lack contextual information related to the query image, resulting in unstable performance. To achieve more robust object detection and enhance the generalization capability of the model, we propose an Oriented Feature Aggregation (OFA) module. This module leverages query feature information to guide the generation of image-specific and more representative prototypes from the support images. Specifically, we first enhance the most salient and important features in the query images through an improved self-attention mechanism. These enhanced features are subsequently used to measure with the support images, generating different weights that facilitate the extraction of support features and the construction of customized prototypes for each query image. Furthermore, we model the fusion of RoI features and class prototypes as an optimal transport problem and improve the loss function to provide more stable matching results for the fusion of features. Extensive experiments on PASCAL VOC and MS COCO demonstrate that our method significantly outperforms previous state-of-the-art methods.

Safety is an unavoidable topic in the field of autonomous driving, and the most crucial aspect of safety in this field is accurate detection of all objects. This article proposes a new multimodal fusion method that combines radar and vision for obstacle detection. The proposed multi-level feature fusion network structure solves the problem of poor detection accuracy caused by incomplete information capture by low-level feature networks. At the same time, the NAM attention module was introduced to further refine the fused features. To solve the problem of low fusion efficiency of heterogeneous sensors, this paper proposes a weight generation method based on spatial attention mechanism (SAF), which achieves efficient fusion of radar and visual features by focusing on key areas in input features and weakening interference areas. In addition, a method of converting radar point clouds into radar images was designed for better network training. The experimental results show that the obstacle detection performance of this method is significantly better than existing technologies in harsh weather conditions (such as rainy and snowy days) and complex lighting conditions (such as nighttime and strong light exposure). Excellent performance was demonstrated on key indicators mAP, AP50, and AP75, demonstrating the high accuracy and strong robustness of the method.

Most of the existing modal fusion methods for multimodal pedestrian detection rely on the traditional neural network architecture, which is limited by the inherent local reduction bias or the quadratic computational complexity, and it cannot fully capture the interaction between the modes. However, the latest research showed that the Mamba architecture based on the state-space model achieves efficient feature extraction with linear computational complexity in long-sequence modeling through a selective scanning mechanism and hardware-aware optimization. Mamba shows higher parameter utilization and linear memory growth characteristics than Transformer in the visual-verbal-temporal cross-modal task. Therefore, we propose a multimodal fusion pedestrian detection scheme based on Mamba. Cross-modal fusion is investigated by associating cross-modal features in a modified Mamba-based hidden state space with a gating mechanism. We design a cross-modality fusion Mamba block to map the modal features of infrared images with those of visible images to a hidden state space for interaction, thus reducing the differences among different modal features. A dual-state-space channel exchange module is designed to promote the fusion of shallow features. A dynamic two-state-spatial fusion module is designed to realize the dynamic deep fusion in the hidden state space. The contribution of different modal features is adaptively adjusted by the gating mechanism to further reduce the modal difference, improving single-mode detection performance. The effectiveness of our model is verified by extensive experiments on four widely used visible-infrared benchmark datasets. The results show that M-fusion has excellent performance in multimodal pedestrian detection.

With the advancement of autonomous driving technology, vehicle trajectory prediction [1] has become one of the important research directions in the field of autonomous driving. Existing vehicle trajectory prediction methods lack hierarchical processing of global semantics and local details, leading to deviations in predicting vehicle driving trajectories in complex road scenarios. Therefore, aiming at the problem that this kind of vehicle trajectory prediction cannot well balance the understanding of the overall scene and the capture of specific details, this paper proposes a vehicle trajectory prediction model based on Hierarchical Spatio-Temporal Interaction Transformer. Firstly, multi-modal feature encoding is completed by extracting features from the processed map semantics and vehicle historical trajectories. Then, hierarchical attention is used to first capture the global features of vehicle groups and road networks, and then focus on the local interaction between single vehicles and surrounding lanes, dynamically fusing to form fusion features with multi-scale constraints. Finally, a decoder is used to generate multi-modal predicted trajectories based on the fusion features, and the optimal trajectory is selected through confidence scoring to complete the trajectory prediction. Experimental results on the Argoverse dataset show that the model achieves the level of other map-based trajectory prediction models in some performance indicators.

Safe and efficient decision-making of autonomous vehicles in complex dynamic scenarios requires a decision-making system with human-like cognitive ability, however, existing deep reinforcement learning methods suffer from insufficient generalization ability of unknown scenarios and exploration-exploitation imbalance. In order to solve the above problems, this study proposes a novel prefrontal cortex (PFC) decision-making model, which builds a multi-module synergistic cognitive architecture by modeling the spatio-temporal reasoning of the lateral prefrontal cortex (LPFC), the reward prediction of the medial prefrontal cortex (MPFC), and the adaptive adjustment function of the anterior cingulate cortex (ACC). The innovations are fusion of graph convolutional network (GCN) and long short-term memory network (LSTM) to capture vehicle interaction features; introduction of unsupervised clustering and deep belief network (DBN) to achieve metacognitive planning of action-reward causal association, proposing a dynamic exploration rate regulation mechanism based on alertness, and realizing strategy optimization in complex scenarios through dopamine-based reward prediction error. In this study, this study test the performance of the method in highway and ring intersection scenarios and compare it with existing deep reinforcement learning (DRL) and graph reinforcement learning (GRL) methods. The experimental results show that the PFC model can perform spatio-temporal and task reasoning, and is able to make better decisions in complex and changing scenarios, which significantly improves the efficiency of access. This study can provide reference value for the development of biological neural models and promote their application in dynamic traffic interaction scenarios.

Continuous-scale trusted safety efficiency evaluation is crucial for the agile development and robust validation of autonomous vehicle intelligence. While the UN R157 Regulation evaluates automated lane-keeping system (ALKS) performance baselines through safe collision plots (SCPs) in various scenario clusters, quantifying the specific ALKS safety efficiency remains challenging. We propose a spectrum quantification approach to evaluate the safety efficiency of autonomous vehicles in cut-in scenarios. First, we collected speed-distance data under different cut-in scenarios and extracted essential spectral features to indicate the vehicle motion parameters during the cut-in process. Second, by utilizing Fourier analysis, a spectral analysis model was built to quantify and analyze the vehicle motion characteristics, providing insights into scenario safety. Finally, we created approximate analytical equations for the normalized disturbance frequencies in the nonlinear response scenarios of autonomous driving systems by combining the SCP with a frequency spectrum analysis model. The results showed that the normalized disturbance frequency in the cut-in scenario was approximately 0.2. When the relative longitudinal distance and speed of the vehicle are the same, if the cut-in speed of the cut-in vehicle is larger, the normalized disturbance frequency is higher, indicating that the cut-in process of the autonomous vehicle is more dangerous and may trigger a collision.

Person Re-identification (ReID) is an important research direction in the field of pattern recognition, which aims to retrieve the same pedestrian in different cameras. The combination of deep learning and attention mechanism greatly improves the accuracy of image retrieval, but previous researchers usually use on-channel or spatial convolution to learn attention, ignoring the connection between attention feature nodes. In this article, we first improve a bottleneck attention module (BAM) to make the learned attention map faster. Secondly, to capture the relevance of each feature node in the global attentional feature map, we design a self-relevant attention module (SRA), which models the global scope structure information and is used to capture the connection between the feature node positions to make the obtained attentional map more robust. Finally, we propose a method to strengthen the attention features, so that the higher attention features around the position also get higher feature values, so that the obtained feature map is more robust. The effectiveness of the model is confirmed in several mainstream pedestrian re-identification datasets, and the proposed model outperforms most state-of-the-art methods.

With the development of productivity and the increase of logistics transport, double-hung trains came into being, but the traffic accidents caused by them are also very serious. Based on vehicle parameters, this paper completed the design of the braking system, and built an A-double model based on Trucksim software to establish the braking system and conduct simulation analysis on typical working conditions. The optimized design results show that the vehicle equipped with fuzzy anti-lock system has the shortest braking distance, and the braking efficiency of A-double is improved compared with the vehicle without anti-lock system or equipped with traditional anti-lock system.

An essential prerequisite for autonomous vehicles deploying in urban scenarios is the ability to accurately recognize the behavioral intentions of pedestrians and other vulnerable road users and take measures to ensure their safety. In this paper, a spatial-temporal feature fusion-based multi-attention network (STFF-MANet) is designed to predict pedestrian crossing intention. Pedestrian information, vehicle information, scene context, and optical flow are extracted from continuous image sequences as feature sources. A lightweight 3D convolutional network is designed to extract temporal features from optical flow. Construct a spatial encoding module to extract the spatial features from the context. Pedestrian motion information are re-encoded using a collection of gated recurrent units. The final network structure is created through ablation research, which introduces attention mechanisms into the network to merge pedestrian motion features and spatio-temporal features. The efficiency of the suggested strategy is demonstrated by comparison experiments on the datasets JAAD and PIE. On the JAAD dataset, the intent recognition accuracy is 9% more accurate than the existing techniques.

Despite the rapid advancement in the field of autonomous driving vehicles, developing a safe and sensible decision-making system remains a challenging problem. The driving decision-making module is one of the most essential sections of the entire autonomous driving system, and the decision generated from it can significantly impinge the lives and property of passengers. Complicated interactions among traffic participants have the most profound impact on the decision-making process, yet the interactions are often simplified or overlooked due to their complexity and implicit nature. To address this issue, this work proposes an Edge-Enhanced Graph Attention Reinforcement Learning (EGARL) framework that aims to make rational driving decisions by comprehensively modeling the interactions among agents. EGARL comprises three core components: a graphical representation of the traffic scenario that covers both topological and interactive information; an Edge-enhanced Graph Attention Network (E-GAT) that utilizes the graphical representation to extract interactive features by comprehensively considering nodes and edges of the graph; and a deep reinforcement learning method that generates driving decisions based on the current state and features extracted from E-GAT. Experimental results demonstrate the satisfying performance of EGARL. Our proposed framework can contribute to the development of intelligent transportation systems, enhancing the safety and efficiency of driving.

Accurate and efficient pedestrian trajectory prediction is one of the key capabilities for the safe operation of self-driving vehicles. Therefore, it is of great significance to study pedestrian trajectory prediction algorithms applicable to complex interaction scenarios. In this study, a spatial gridding-based multi-attention generative adversarial network (SGMA-GAN) is proposed, which is modeled with generative adversarial network as the main framework. Firstly, the map information is gridded to better represent the pedestrian state information in tensor form, improve the stability of the state space and network structure. Secondly, temporal and spatial attention mechanisms are introduced to account for the effects of historical trajectories and spatial interaction features. Finally, the model is evaluated with both Eidgenössische Technische Hochschule (ETH) and University of Cyprus (UCY) datasets. The results showed that as the prediction step size gradually increased, compared with the relatively new SGANv2, the mean average displacement error (ADE) and Final displacement error (FDE) of SGMA-GAN in five scenarios increased by 10.61

At present, autonomous driving decision-making solutions take few elements into account while ignoring the unpredictable nature of driving behavior, which makes it challenging to manage complicated traffic situations. To this end, we present a decision-making architecture in this paper that enhances the existing reinforcement learning methodology by combining the bootstrapped technique and the random prior network (RPN). The RPN can give each learner a neural network with unique weights to avoid the contingency created by the artificially built prior functions, while the Bootstrapped technique can balance out the exploration and exploitation. The ego vehicle was trained by three algorithms and verified in random environments to evaluate the effectiveness of our method. The results show that our algorithm outperformed the current reinforcement learning algorithms.

Background: Pedestrian trajectory prediction plays a crucial role in ensuring the safe and efficient operation of autonomous vehicles in urban environments. As autonomous driving technology continues to advance, accurate anticipation of pedestrians' motion trajectories has become increasingly important for informing subsequent decision-making processes. Pedestrians are dynamic and unpredictable agents, and their movements can vary greatly depending on factors, such as their intentions, interactions with other pedestrians or vehicles, and the surrounding environment. Therefore, developing effective methods to predict pedestrian trajectories is essential to enable autonomous vehicles to navigate and interact with pedestrians in a safe and socially acceptable manner. Various methods, both patented and non-patented, have been proposed, including physics-based and probability- based models, to capture the regularities in pedestrian motion and make accurate predictions. Objective: This paper proposes a pedestrian trajectory prediction method that combines a Gaussian mixture model and an artificial potential field. Methods: The study begins with an analysis of pedestrian motion patterns, allowing for the identification of distinct patterns and incorporating speed as an influential factor in pedestrian interactions. Next, a Gaussian mixture model is utilized to model and train the trajectories of pedestrians within each motion pattern cluster, effectively capturing their statistical characteristics. The trained model is then used with a regression algorithm to predict future pedestrian trajectories based on their past positions. To enhance the accuracy and safety of the predicted trajectories, an artificial potential field analysis is employed, considering factors such as collision avoidance and interactions with other entities. By combining the Gaussian mixture model and artificial potential field, this method provides an innovative and patentable approach to pedestrian trajectory prediction. Results: Experimental results on the ETH and UCY datasets demonstrate that the proposed method combining the Gaussian mixture model and artificial potential field outperforms traditional Linear and social force models in terms of prediction accuracy. The method effectively improves accuracy while ensuring collision avoidance. Conclusion: The proposed method combining a Gaussian mixture model and an artificial potential field enhances pedestrian trajectory prediction. It successfully captures the differences between pedestrians and incorporates speed, improving prediction accuracy.

Currently, in the process of autonomous parking, the algorithm detection accuracy and rate of parking spaces are low due to the diversity of parking scenes, changes in lighting conditions, and other unfavorable factors. An improved algorithm based on YOLOv5-OBB is proposed to reduce the computational effort of the model and increase the speed of model detection. Firstly, the backbone module is optimized, the Focus module and SSP (Selective Spatial Perception) module are replaced with the general convolution and SSPF (Selective Search Proposals Fusion) modules, and the GELU activation function is introduced to reduce the number of model parameters and enhance model learning. Secondly, the RFB (Receptive Field Block) module is added to fuse different feature modules and increase the perceptual field to optimize the small target detection. After that, the CA (coordinate attention) mechanism is introduced to enhance the feature representation capability. Finally, the post-processing is optimized using spatial location correlation to improve the accuracy of the vehicle position and bank angle detection. The implementation results show that by using the improved method proposed in this paper, the FPS of the model is improved by 2.87, algorithm size is reduced by 1 M, and the mAP is improved by 8.4% on the homemade dataset compared with the original algorithm. The improved model meets the requirements of perceived accuracy and speed of parking spaces in autonomous parking.

在森林火灾早期,烟雾形态多变、火焰较小、背景噪声干扰等使现有目标检测算法在应对复杂森林火灾环境时,表现出精度低、误检漏检和速度慢.针对这些问题,提出基于改进的YOLOX森林火灾检测模型.在YOLOX的Neck网络中添加了SimAM注意力模块,增强算法对烟雾和火焰的特征提取能力,抑制背景噪声,并对Backbone进行优化,加快网络的推理速度;考虑到烟雾与火焰在空间上的相关性,改进了YOLOX的后处理算法,进一步提高了目标检测的精度.实验结果表明,与原YOLOX算法相比,改进算法在自制森林火灾数据集上mAP提高了 2.4%,对烟雾和火焰的识别率明显提升.

To explore the influence of path deflection on crack propagation, a path planning algorithm is presented to calculate the crack growth length. The fatigue crack growth life of metal matrix composites（MMCs） is estimated based on an improved Paris formula. Considering the different expansion coefficient of different materials, the unequal shrinkage will lead to residual stress when the composite is molded and cooled. The crack growth model is improved by the modified stress ratio based on residual stress. The Dijkstra algorithm is introduced to avoid the cracks passing through the strengthening base and the characteristics of crack steps. This model can be extended to predict crack growth length for other similarly-structured composite materials. The shortest path of crack growth is simulated by using path planning algorithm, and the fatigue life of composites is calculated based on the shortest path and improved model.And the residual stress caused by temperature change is considered to improve the fatigue crack growth model in the material. The improved model can well predict the fatigue life curve of composites. By analyzing the fatigue life of composites, it is found that there is a certain regularity based on metal materials, and the new fatigue prediction model can also reflect this regularity.