ViFusion: Enhanced 3D object detection via virtual point augmentation and dynamic multi-source feature fusion

This article introduces a multilayer navigation algorithm for a fleet of unmanned aerial vehicles (UAVs). The proposed architecture consists of a fusion of virtual point controllers and potential field techniques. On the one hand, a potential function is constructed for every agent such that its position smoothly and robustly converges to a virtual guidance point while avoiding collisions with other agents. The virtual points, on the other hand, are controlled to fulfill a swarm control goal such as target tracking, station keeping, or search and rescue missions. Therefore, the suggested system has two levels of hierarchy, but the algorithm can be generalized for multiple levels. The vehicle translational and rotational dynamics are controlled using an internal loop based on gradient tracking and sliding mode controllers. The architecture is validated in simulations and real-time experiments, showing good performance for the closed-loop system.

Advanced driver assistance systems (ADAS) mainly consist of three components: environmental perception, decision planning, and motion control. As a fundamental component of the ADAS environmental perception system, 3D object detection enables vehicles to avoid obstacles and ensure driving safety only through accurate and real‐time prediction and localization of three‐dimensional targets such as vehicles and pedestrians in road scenes. Therefore, to improve both the real‐time performance and accuracy of 3D object detection, we propose a lightweight depth prediction‐based 3D object detection model with multi‐scale fusion—SPWS‐Transformer. First, to enhance the model's accuracy, we propose a feature extraction network incorporating multi‐scale feature fusion and depth prediction. By designing a multi‐scale feature fusion module, we effectively combine multi‐scale semantic and fine‐grained information from feature maps of different scales to enhance the network's feature extraction capability. To capture spatial information from the feature maps, we apply convolution, group normalization, and nonlinear activation operations on the fused feature maps to generate depth feature maps. Both the fused feature maps and depth feature maps serve as inputs for subsequent network stages. To further improve accuracy, we leverage the long‐range modelling advantages of Transformers by designing a feature enhancement encoder to strengthen the representation capability of depth feature maps. We incorporate a dilated encoder to perform positional encoding on depth feature maps and utilize multi‐head self‐attention mechanisms to capture contextual relationships within the input scene, thereby enhancing the detection capability of the 3D object detection network. Then, to improve real‐time performance, we design a decoder structure with scale‐aware attention. By predefining masks of different scales, we adaptively learn a scale‐aware filter using depth and visual features to enhance object queries. Finally, on the KITTI dataset, the improved algorithm achieves an AP of 24.66% for the car category, with more significant improvements in detection accuracy under the ‘hard’ difficulty level. The model achieves an inference time of 24 ms.

With the increasing demand for object detection accuracy in scenarios such as intelligent transportation and autonomous driving, methods that utilize point cloud and voxel features to achieve multi-modal feature fusion have become increasingly common in 3D object detection. However, existing methods often rely on inefficient linear fusion strategies during the multi-modal feature fusion process, which fails to adequately capture the dependencies between multi-source features, leading to insufficient feature integration. Additionally, during feature extraction, limitations in network architecture result in a lack of interaction between shallow and deep features, causing the loss of fine-grained feature information, which particularly affects the detection of small objects.To address these issues, we propose the Hybrid Point-Voxel-Transformer Network HPP-TNet , a two-stage object detection framework that integrates point and pillar features. Specifically, we design a fine-grained pillar feature extraction module CFPEM , which effectively alleviates the feature loss problem caused by voxel downsampling through shallow-deep feature interaction and lightweight attention design. Next, we develop a transformer-based multi-scale feature fusion module TMFFM , which dynamically achieves cross-modal associations through amulti-head attention mechanism, enhancing context-aware features and fully realizing multi-source feature fusion. Experiments on the KITTI dataset demonstrate that our proposed algorithm achieves competitive detection performance compared to several state-of-the-art methods, particularly in the Cyclist and Pedestrian categories. Our code will be open-sourced soon.

With the improvement of the azimuth width and azimuth resolution of the spaceborne SAR, the azimuth scan angle range of the antenna beam increases rapidly, and the beam center deviates significantly from the virtual rotation point under the control of uniform scanning, and the traditional virtual rotation point model becomes invalid. At this time, it is impossible to continue to use the virtual rotation point to perform azimuth deramping, and it is difficult to continue to apply the frequency domain imaging algorithm based on azimuth deramping. This paper proposes an improved azimuth deramping technology that does not rely on virtual rotation points. This technology can accurately remove Doppler aliasing caused by beam azimuth scanning and solve the problem of virtual rotation point model failure in sliding spotlight mode with large azimuth scanning angles.

We propose a novel formation control method considering the challenges of formation keeping, unmanned aerial vehicle (UAV) control, and position adjustment in UAV Base Station (UAV-BS) formation control scenarios. The method is based on a loose coupling structure inspired by the traditional virtual structure method. Firstly, a type of new formation is designed for UAV-BS group. When the UAV-BS group keep a stable network, the centers of all UAV-BS moving trajectory, e.g., circular, are treated as virtual points to form a rigid body. When the UAV-BS group transforms between different formations, the real positions of the UAV-BS group are used to form a rigid body. Secondly, we redesign a trajectory tracking method for UAV-BS to control the flight direction. A flight direction controller is designed for UAV-BS, based on nonlinear guidance logic. According to the empirical model of error feedback, an error corrector is designed to control the position error between the UAV-BS and the virtual point. Simulation results show that the proposed algorithm can satisfactorily complete the formation and control of UA-BS. Besides, the algorithm can reduce the difficulty of formation control of UAV-BS and the requirements of formation control for computing power and network communication.

Segmentation of vessels in biomedical images is important as it can provide insight into analysis of vascular morphology, topology and is required for kinetic analysis of flow velocity and vessel permeability. Intravital microscopy is a powerful tool as it enables in vivo imaging of both vasculature and circulating cells. However, the analysis of vasculature in those images is difficult due to the presence of cells and their image gradient. In this paper, we provide a novel method of segmenting vessels with a high level of cell related clutter. A set of virtual point pairs ("vessel probes") are moved reacting to forces including Vessel Vector Flow (VVF) and Vessel Boundary Vector Flow (VBVF) forces. Incorporating the cell detection, the VVF force attracts the probes toward the vessel, while the VBVF force attracts the virtual points of the probes to localize the vessel boundary without being distracted by the image features of the cells. The vessel probes are moved according to Newtonian Physics reacting to the net of forces applied on them. We demonstrate the results on a set of five real in vivo images of liver vasculature cluttered by white blood cells. When compared against the ground truth prepared by the technician, the Root Mean Squared Error (RMSE) of segmentation with VVF and VBVF was 55% lower than the method without VVF and VBVF.

To address the practical engineering challenges of limited fault samples for high-voltage circuit breaker spring operating mechanisms and the inability of single features to fully reflect equipment status, this paper proposes a small-sample fault diagnosis method based on multi-source feature fusion and Stacking ensemble learning. First, a multi-source sensing system containing MEMS (Micro-Electro-Mechanical System) pressure and travel, coil, and motor current was constructed to achieve comprehensive monitoring of the mechanical and electrical states of a 220 kV circuit breaker; in particular, the introduction of non-invasive MEMS sensors effectively solves the difficulty of capturing static spring fatigue characteristics inherent in traditional methods. Second, a high-dimensional feature space was constructed using Savitzky-Golay filtering and physical feature extraction techniques. To address the characteristics of small-sample data distribution, a two-layer Stacking ensemble learning model based on 5-fold cross-validation was designed. This model utilizes the SVM (Support Vector Machine), RF (Random Forest), and KNN (K-Nearest Neighbors) as base classifiers and Logistic Regression as the meta-learner, achieving an adaptive fusion of the advantages of heterogeneous algorithms. True-type experimental results show that the average diagnostic accuracy of this method under normal conditions and four typical fault conditions reaches 96.1%, which is superior to single base models (the RF was 94.2%). Feature importance analysis further confirms that closing and opening pressures are the most critical features for distinguishing mechanical faults. This study provides effective theoretical basis and technical support for condition-based maintenance of high-voltage circuit breakers under small-sample conditions.

3D object detection based on sparse convolution neural network and feature fusion for autonomous driving in smart cities

Point cloud-based three-dimensional (3D) object detection is a critical task in autonomous driving, robotics, and augmented reality, where accurate localization and classification of objects are essential under diverse and challenging scenarios. This work introduces a point cloud-based multi-target 3D object detection framework using Light Detection and Ranging (LiDAR) sensors. The key contribution lies in improving the encoding technique to preserve spatial information and adapt to varying point densities, enabling efficient processing of raw LiDAR data for robust 3D detection. Specifically, we design an efficient network architecture that incorporates a single unified classification head to jointly handle positive and negative samples, simplifying the design and improving training stability. Our model detects all classes (car, pedestrian, cyclist) simultaneously within a single network, enhancing computational efficiency while avoiding separate networks for each class. Furthermore, the proposed feature extraction pipeline combines a voxel-based encoder with a sparsely embedded convolutional backbone and a feature pyramid network, facilitating multi-scale feature representation and effective detection of objects at different scales. This encoding method facilitates the efficient processing of raw LiDAR data, enabling accurate object detection across diverse scenarios. As a result, our model achieves state-of-the-art performance on the KITTI dataset, surpassing baseline methods such as PointPillars and VoxelNet by delivering superior Average Precision (AP) across all difficulty levels for the car class in both bird’s-eye view (BEV) and 3D object detection tasks. Specifically, it achieves AP40 scores of 92.84%, 88.63%, and 85.68% in BEV detection and 87.84%, 76.65%, and 73.60% in 3D detection for Easy, Moderate, and Hard levels, respectively. These results highlight the effectiveness of our encoding technique in enhancing model efficiency and detection accuracy across diverse scenarios.

Positioning services are increasingly used for applications such as navigation, advertising and social media. While outdoor navigation based on GPS and/or cellular systems works well, indoor navigation is a much tougher challenge. This paper proposes an indoor localization using Wi-Fi received signal strength indicator (RSSI) measurements from existing access points. Using RSSI data from multiple access points, additional access points will be added in the system thru linear regression statistical model to create virtual access points (VAPs). This paper fuses the low-cost and flexible system of fingerprint localization with VAP creation and the improvement in accuracy by adapting Kalman Filter (KF) and Particle Filter (Pf). The effectiveness of using KF with VAP has demonstrated positive results in previous research. This paper aims to integrate the use of PF in the system. To improve the performance of indoor localization, the proposed model will be an improved virtual access point localization, proposed filtering (VAP+KF+PF).

Aiming at the local minimum problem and target unreachable problems in the path planning of unmanned surface vehicle (USV), a path planning algorithm of USV considering virtual target point is proposed. For the target unreachable problem, a repulsive force potential field function is created. According to the measured distance between the target and the obstacle, the repulsive force of the target is zero, so that the USV can reach the target point. For the local minimum problem, the local minimum caused by various obstacles is analyzed, simulated annealing (SA) and artificial potential field approach (APFA) are combined to solve the minimum point problem caused by general obstacles. For the local minimum problem caused by special U-shaped obstacles, a virtual target point algorithm (VTPA) is established to solve this problem. The simulation results prove that this algorithm solves the problem of long path caused by too large repulsive force. Compared with Dual-Tree Rapidly exploring Random Tree (DT-RRT) algorithm, the efficiency of this algorithm has been significantly improved. Algorithm running time was reduced by 33.4%. Path length was reduced by 17.8%. Not only the time is saved, but also the path planning is optimized, so that the speed of the algorithm is accelerated.

<div class="section abstract"><div class="htmlview paragraph">3D object detection based on the point cloud is a critical technology for environmental perception in autonomous driving systems. However, it has long faced dual challenges of accuracy and efficiency due to the sparsity and irregularity of point cloud data and the complicated driving scenarios. This paper proposes a novel multi-frame fusion 3D object detection algorithm, MFFormer3D, embedding object queries. The method first removes ground points from the raw point cloud to reduce subsequent computational load. It then employs Spatially Grouped Sparse Convolution Feature Extraction Layer (SGSC-FEL) to optimize the backbone network for feature extraction. Subsequently, we introduce a Cross-frame Object Query Filter (COQF) mechanism for efficient multi-frame fusion and a Motion Correction Module (MCM) to compensate for ego-motion effects. Experiments on the nuScenes dataset demonstrate that MFFormer3D outperforms the baseline method by 2.25% in NDS and 7.73% in mAP while maintaining an inference speed of 11.1 FPS. Ablation studies further validate the effectiveness of each module.</div></div>

Directly estimating the position and pose of objects in a 3D coordinate system from monocular images is an extremely challenging task. Many existing methods typically rely on direct regression or geometric projection techniques to obtain a single depth value for the target, but these approaches often struggle to provide sufficiently accurate depth estimates in complex scenes. To address this issue, we propose an improved 3D object detection method that enhances depth estimation accuracy by fusing depth information from multiple keypoints. Specifically, the method first predicts multiple corner points and surface center points of the 3D bounding box. Then, based on the image heights and physical heights of these keypoints, multiple depth values are computed through geometric projection. These depth values are subsequently fused with the depth information directly regressed by the depth branch to achieve more reliable depth estimation. Additionally, to improve the model's capability in detecting objects of varying scales, we introduce a multi-scale attention enhancement module into the feature extraction network. Finally, experimental results on the KITTI dataset demonstrate that the proposed method significantly improves the performance of 3D object detection.

This paper designs a broadcast and television content recommendation algorithm (DFFRA) based on deep feature fusion. This algorithm employs a three-layer cascade architecture to achieve precise recommendations through feature extraction, dynamic fusion, and interest prediction. The feature extraction layer processes multimodal content features and user behavior sequences in parallel. The feature fusion layer innovatively designs an attention-gated fusion unit. The recommendation prediction layer outputs user interest scores. Experiments on the TVRec-2024 dataset show that, when K=10, DFFRA achieves an accuracy of 82.3%, a recall of 78.5%, an F1 score of 80.3%, and a MAE of 0.62. These improvements are 4.2%, 3.8%, 4.0%, and 0.08, respectively, compared to the next-best algorithm, AttnRec, with the lowest standard deviation (<0.7). Compared to the traditional CF algorithm, DFFRA achieves a 15.6% improvement in accuracy and 8.9% and 6.5% improvement in F1 score over CNNRec and LSTMRec, respectively. Stability tests show that when the training data is reduced by 50%, the F1 score of DFFRA only decreases by 3.2%. Training time increases linearly with the number of users, validating its advantages in data-sparse and large-scale user scenarios.

The canopy temperature of winter jujube serves as a direct indicator of plant water status and transpiration efficiency, making its accurate prediction a critical prerequisite for effective water management and optimized growth conditions in greenhouse environments. This study developed a data-driven model to forecast canopy temperature. The model serially integrates a Long Short-Term Memory (LSTM) network and a Random Forest (RF) algorithm, leveraging their complementary strengths in capturing temporal dependencies and robust nonlinear fitting. A three-stage framework comprising temporal feature extraction, multi-source feature fusion, and direct prediction was implemented to enable reliable nowcasting. Data acquisition and preprocessing were tailored to the greenhouse environment, involving multi-sensor data and thermal imagery processed with Robust Principal Component Analysis (RPCA) for dimensionality reduction. Key environmental variables were selected through Spearman correlation analysis. Experimental results demonstrated that the proposed LSTM–RF model achieved superior performance, with a determination coefficient (R2) of 0.974, mean absolute error (MAE) of 0.844 °C, and root mean square error (RMSE) of 1.155 °C, outperforming benchmark models including standalone LSTM, RF, Transformer, and TimesNet. SHAP (SHapley Additive exPlanations)-based interpretability analysis further quantified the influence of key factors, including the “thermodynamic state of air” driver group and latent temporal features, offering actionable insights for irrigation management. The model establishes a reliable, interpretable foundation for real-time water stress monitoring and precision irrigation control in protected winter jujube production systems.