🔬 Research Overview
In recent years, the laboratory has mainly engaged in research and development in IoT and Edge Intelligence, machine vision and 3D spatial understanding, natural language processing and intelligent documents, and embodied intelligence and multi-agent systems. We have undertaken more than 10 national and provincial vertical scientific research projects, including those funded by the National Natural Science Foundation of China and provincial natural science foundations, published more than 100 papers in domestic and international journals and international conferences, and obtained more than 30 authorized patents. We have actively explored new models of industry-education-research-application through Corporate Partnership and completed more than 20 university-industry cooperation projects.
🎯 Research Directions
📡IoT and Edge Intelligence
Focusing on the computing architecture and collaboration mechanisms of edge intelligence in IoT environments, this explores the deployment and inference optimization of lightweight models on resource-constrained devices. The key research areas include device-edge-cloud collaborative learning, federated learning, and distributed model update strategies, addressing critical issues such as low latency, high energy efficiency, and data privacy protection.
🤖Machine vision and 3D spatial understanding
This work is dedicated to recovering the dense three-dimensional structure of scenes from monocular or multi-view images, and combines deep learning to achieve joint semantic and geometric understanding. It focuses on breakthroughs in three-dimensional reconstruction in dynamic environments, pose estimation of weakly textured objects, and cross-view scene re-identification, providing a unified visual representation for object detection, semantic segmentation, and spatial relationship reasoning.
📝Natural Language Processing and Intelligent Documents
Focusing on intelligent processing of unstructured and semi-structured documents, explore core technologies such as document layout analysis, table understanding, information extraction, and visual question answering. Emphasize research on multimodal document pretraining models, knowledge extraction under weak supervision, and document-level semantic reasoning, to achieve fully automated understanding and structured output for multi-source documents ranging from scanned copies and PDFs to web pages.
🔌Embodied intelligence and multi-agent systems
Focusing on the perception-decision-execution loop of embodied agents in physical environments, this research integrates multi-agent reinforcement learning to explore mechanisms of group collaboration and competition. It focuses on key issues such as hierarchical decision-making in partially observable environments, intent understanding and task allocation in human-machine collaboration, and communication and credit assignment strategies among multiple agents.