LiDAR has become a vital sensor for autonomous driving scene understanding. To meet the accuracy and speed of LiDAR point clouds semantic segmentation, an efficient model ACPNet is proposed in this paper. In the feature extraction stage, the backbone is constructed with asymmetric convolutions, so the skeleton of the square convolution kernel is enhanced, which leads to greater robustness to target rotation. Moreover, a contextual feature enhancement module is designed to extract richer contextual features. During training, global scaling and global translation are performed to enrich the diversity of datasets. Compared with the baseline network PolarNet, the mIoU of ACPNet on the SemanticKITTI, SemanticPOSS and nuScenes datasets are improved by 5.1%, 1.6% and 2.9%, respectively. Meanwhile, the speed of ACPNet is 14 FPS, which basically meets the real-time requirements in autonomous driving scenarios. The experimental results show that ACPNet significantly improves the performance of LiDAR point cloud semantic segmentation.

Fusing LiDAR point cloud and camera image for 3D object detection in autonomous driving has emerged as a captivating research avenue. The core challenge of multimodal fusion is how to seamlessly fuse 3D LiDAR point cloud with 2D camera image. Although current approaches exhibit promising results, they often rely solely on fusion at either the data level, feature level, or object level, and there is still a room for improvement in the utilization of multimodal information. We present an advanced and effective multimodal fusion framework called EPAWFusion for fusing 3D point cloud and 2D camera image at both data level and feature level. EPAWFusion model consists of three key modules: a point enhanced module based on semantic segmentation for data-level fusion, an adaptive weight allocation module for feature-level fusion, and a detector based on 3D sparse convolution. The semantic information of the 2D image is extracted using semantic segmentation, and the calibration matrix is used to establish the point-pixel correspondence. The semantic information and distance information are then attached to the point cloud to achieve data-level fusion. The geometry features of enhanced point cloud are extracted by voxel encoding, and the texture features of image are obtained using a pretrained 2D CNN. Feature-level fusion is achieved via the adaptive weight allocation module. The fused features are fed into a 3D sparse convolution-based detector to obtain the accurate 3D objects. Experiment results demonstrate that EPAWFusion outperforms the baseline network MVXNet on the KITTI dataset for 3D detection of cars, pedestrians, and cyclists by 5.81%, 6.97%, and 3.88%. Additionally, EPAWFusion performs well for single-vehicle-side 3D object detection based on the experimental findings on DAIR-V2X dataset and the inference frame rate of our proposed model reaches 11.1 FPS. The two-layer level fusion of EPAWFusion significantly enhances the performance of multimodal 3D object detection.

China is a large agricultural country, and in the process of agricultural production, it is very important to make accurate prediction of soil moisture. To address the problems of local minimization and slow convergence of traditional BP (back propagation) neural network in the prediction process, this paper combines LSTM (long short-term memory) and Elman neural network with traditional BP neural network model, and proposes a method based on LSTM and Elman neural network for soil moisture prediction. A soil moisture prediction method based on LSTM and Elman neural network is proposed. The prediction model of LSTM and Elman neural network was developed, and the soil moisture of Xilinguole grassland in Inner Mongolia was predicted and experimented. The results show that the accuracy of the model is higher than that of the unoptimized BP neural network. The model is able to reduce the use of moisture sensors significantly, which reduces the cost for agricultural production.

Robotaxis are emerging as a key component of urban transportation. However, most current robotaxis rely on preset pick-up points, which cluster robotaxi activity and intensify local traffic congestion, increase deadheading and detour distances, extend roadway occupancy, and raise operating costs. GNSS-based localization may be unreliable especially in urban area. Generally, human drivers can communicate with passengers through text/voice interactions to accurately locate and reach the human’s location. How to align the robotaxi’s visual recognition results with human text/voice description is the key challenge. Therefore, we propose a comprehensive framework, including its architecture, requirment, opertional logic, workflow, to enhance the passenger-robotaxi interaction for accurate passenger pick-up without preset pick-up points. In this framework, the key algorithm is VLMIdentification, a realtime human identification model based on LVLM. VLMIdentification comprises three modules: i) Human Input Processing, which extracts textual features from passenger text/voice and converts human-centric descriptions into third-person, robotaxi-centric attributes; ii)Candidate Searching, which couples conventional detectors with an LVLM to adapt detection thresholds to scene complexity and to translate detections into textual descriptors; and iii) Human Identification, which performs matching between the processed passenger description and candidates to locate the correct human. We defined the human identification with multimodal task, proposed evaluation metrics, and built a new HID (Human Identification with Description) dataset based on existing autonomous driving datasets. Experimental results demonstrate that VLMIdentification outperforms baseline in comprehensive metrics and retains robust, performance under adverse environment and cross-scenario generalization tests, thereby confirming its generalization and robustness. Code is available at https://github.com/fanwu66/VLMIdentification

Dynamic spectrum reallocation must mitigate severe link degradation under limited search and frequency switching. We study a risk-triggered controller combining receiver/aggressor ranking, screened greedy repair and conditional two-coordinate refinement under cumulative co-channel and adjacent channel interference. With four frequency candidates per link and a 12-link greedy screen, each noninitial decision requires at most 104 physical evaluations and respects a hard switching limit. Component, disturbance recovery and single-factor experiments characterize the operating trade-offs. Findings from these studies motivate a separately frozen tail risk validation using 30 new seeds at each of two network sizes, yielding 240 closed-loop trajectories. Its sole primary endpoint is the trajectory-mean cross-link conditional value-at-risk of signal-to-interference-plus-noise ratio (SINR) shortfall. In the prescribed 30-link scenarios, the controller reduces this endpoint by 10.61% against ranked coordinate search and 11.97% against a literature-informed memetic reference. Both comparisons survive Holm correction. An advantage over the no-refinement ablation is not established, nor is any primary comparison at 14 links. All methods fail to attain fully feasible slots in the 30-link validation. A separate screening size sensitivity study finds lower 30-link tail risk with a 24-link screen under the same evaluation ceiling; the original 12-link choice is therefore not optimal among the tested settings. Post-review rescorings of the R4 decisions retain the 30-link comparator advantages for raw dB shortfall but not the absolute linear SINR shortfall. Thus, the observed ranking depends on the declared loss definition. The evidence therefore supports conditional mitigation of severe link degradation under that definition, not metric-independent superiority, service-level compliance or an end-to-end real-time guarantee.

To address the critical requirements for electromagnetic spectrum orchestration in complex ultra-dense communication environments, this paper proposes an Expert Knowledge-Guided improved NSGA-III framework to solve large-scale frequency assignment problems efficiently. which is built upon the standard NSGA-III architecture as the algorithmic backbone. Traditional multi-objective evolutionary algorithms often struggle with slow convergence and insufficient local search capabilities when navigating high-dimensional, strongly constrained search spaces. In this study, we first introduce a Conflict Graph-based Intelligent Initialization strategy to generate high-quality initial populations by constructing an interference conflict graph based on network topology. Second, a Knowledge-Guided Mutation operator is designed to precisely identify and reconfigure conflicting communication nodes using physical layer indicators. Furthermore, a Best-individual Guided Double-Scale Mutation mechanism is incorporated to dynamically balance global exploration and local exploitation. Experimental results on complex multi-node datasets demonstrate that EKG-NSGA-III significantly outperforms the standard NSGA-III and other baseline algorithms in terms of Hypervolume and Inverted Generational Distance. Specifically, for the 200 nodes scenario, the proposed method achieves a 25.7% improvement in IGD and a 4.2% increase in HV compared to the standard NSGA-III. The proposed algorithm provides a robust and efficient solution for spectrum management in complex urban electromagnetic environments, such as future smart city infrastructures.