Finished Works:

1. We integrate multi-platform and multi-modal data sources to construct a unified object detection and tracking processing framework, realizing information fusion and enhanced perception.
2. For different platform resource constraints, we implement and deploy mainstream detection algorithms such as RTMDet and Deformable DETR, balancing detection accuracy with computational efficiency.
3. We implement advanced tracking algorithms like DeepSORT and ByteTrack to achieve stable target association and trajectory maintenance in complex scenarios.
4. Algorithm optimization and deployment tests are conducted on various computing platforms to verify the system’s real-time performance, robustness, and collaborative perception capabilities across multiple scenarios.

Finished Works:

1. We conducted in-depth research into mainstream methods in 3D human pose estimation, ultimately selecting MotionBERT as our experimental model. We thoroughly understood its pre-training + fine-tuning framework and its DSTformer encoder structure.
2. We reproduced the MotionBERT method on the Human3.6M dataset, implementing the model’s training and testing pipeline, including data preprocessing, network architecture configuration, and training parameter settings.
3. The experimental results of MotionBERT were compared against mainstream methods such as VideoPose3D, UGCN, and PoseFormer, using MPJPE (Mean Per Joint Position Error) as the evaluation metric to validate its superiority.
4. Finally, we visualized and analyzed the experimental results, highlighting MotionBERT’s advantages in terms of pose estimation accuracy and robustness, while also discussing its limitations and potential future optimization directions.

Finished Works:

1. Developed virtual radar subsystem models covering signal generation, echo simulation, clutter/jamming, and anti-jamming processing across air-to-air, air-to-ground, and air-to-sea modes. Integrated mainstream anti-jamming algorithms (STAP, sidelobe cancellation) and provided real-time signal visualization and effect evaluation.
2. Designed an interactive human-machine interface for combat drills and training, enabling human-in-the-loop simulation control and flexible mission configuration for interventional teaching.
3. Developed and validated a target tracking and data fusion module, creating configurable aerial target motion models for accurate echo simulation. Integrated and optimized typical tracking algorithms (Kalman filter, JPDA) for robust multi-target tracking. Developed a cooperative data fusion module for multiple radar platforms using Bayesian estimation and neural networks, significantly boosting tracking robustness and evaluation in complex environments.

Finished Works:

1. We designed an end-to-end Few-Shot Learning classification model based on meta-learning principles. This significantly enhances the model’s generalization capability in scenarios with extremely limited samples.
2. We utilized 3D Convolutional Neural Networks (3D-CNN) to extract both spatial and spectral features from hyperspectral images. This effectively uncovers the inter-dimensional relationships within the data.
3. We incorporated a deep Relation Network that uses 2D convolutional modules to compute relational scores between samples. This enables pixel-wise class prediction, boosting discriminative performance under few-shot conditions.

Finished Works:

1. Multi-dimensional Assessment with AHP: We designed a multi-dimensional global equity assessment system using the Analytic Hierarchy Process (AHP). This resulted in the Global Equity Index (GIEI), providing a comprehensive score for each country.
2. National Classification via K-means++: We applied the K-means++ clustering algorithm to categorize countries. This helped us uncover distribution features and cluster patterns across different equity levels.
3. Coupling Economic Growth with Resource Development: Our model integrates resource development factors, like mining revenue, into an economic growth framework. This allowed us to build a coupling model of equity and economic interests.
4. Sensitivity Analysis for Robustness: We conducted sensitivity tests on key weights and parameters within the model. This evaluated how variable changes impact the GIEI and classification results, thereby enhancing the model’s robustness and interpretability.

Finished Works:

1. Systematic Evaluation of Mainstream Architectures: We conducted a comprehensive evaluation of three representative semantic segmentation architectures: FCN (with HRNet backbone), DeepLabV3+ (with ResNet backbone), and PSPNet (with ResNet backbone). Our experiments spanned three typical remote sensing datasets (LoveDA, Potsdam, and Vaihingen), involving 27 model training and evaluation tasks to thoroughly compare their performance in remote sensing scenarios.
2. Analysis of Architectural Adaptability: We systematically analyzed the adaptability differences of these semantic segmentation architectures in remote sensing images. This revealed their performance limitations in complex backgrounds and multi-scale object scenarios, providing a reliable benchmark assessment and model selection reference for few-shot semantic segmentation tasks in marine remote sensing.
3. Preliminary Framework for Cross-Domain Transfer: We established a preliminary framework for model structure analysis and adaptability specifically designed for transferring from general remote sensing to marine remote sensing tasks. This offers a systematic theoretical foundation and practical support for future work in key areas such as transfer learning, structural fine-tuning, and lightweight deployment.