Research

Publication

Research

LLM

• LLM Agents Consensus for AD Simulation | RA        Oct 2023 - Feb 2024

  • Advisor: Prof. Dequan Wang

  • Integrated the Raft algorithm with LLMs to achieve consensus among LLM Agents in AD simulation.

  • Proposed a new method combining Raft and LLM has been proposed, dynamically grouping agents to achieve information synchronization, ultimately ensuring communication among LLMs and environmental stability during the simLLM simulation process.

  • Constructed an AD simulation env that synchronizes and merges content among LLM agents, proving Raft's importance. The first-author paper has been submitted to ICIR 2024 Workshop AGI.

Graphic

• Image Generation and Editing System based on Diffusion Models |RA        May 2023 - Present

  • Advisor: Prof. Yichao Yan

  • Proposed a method combining Gaussian splatting and morphing field reconstruction to achieve video obstacle and shadow removal.

  • Integrates the concepts of dynamic scene reconstruction and HyperNeRF, taking into account the capabilities of Gaussian splatting, suitable for more powerful video reconstruction and shadow obstacle removal work.

  • Boosted the PSNR by 3-4% compared to traditional methods. The relevant paper is being prepared for submission to ECCV 2024

Vision

• Small Object Detection and Recognition in Autonomous Driving | RA        March 2023 - Feb 2024

  • Advisor: Prof. Manhua Liu

  • Proposed adjustments to the DETR network for small object detection tasks and explored the effectiveness of mask networks in detecting small objects.

  • Combined and adjusted DINO, DETR, and other end-to-end object detection networks with mask self-supervised networks for enhanced small object detection, exploring mask network capabilities in this context.

  • Boosted DETR by 3 to 5 AP points, enhancing small object detection and proving mask networks’ effectiveness for such tasks.

• Research on Active Target Tracking Algorithms in CV | RA        June 2022 - August 2023

  • Advisor: Prof. Yu Qiao

  • Proposed a target detection comparison environment based on active object tracking, comparing and evaluating mainstream video object tracking algorithms.

  • Integrated UnrealCV to build an object tracking and detection system in UE4, allowing for evaluation and comparison of different algorithms to derive final algorithm assessment outcomes.

  • Built a simulated environment with a virtual reality setup and integrated systems, setup and integrated systems, supporting multi-environment object tracking using various datasets. All environment information can be viewed here.

• Survey of Computer Vision Recognition Algorithms | Intern        June 2022 - Sep 2022

  • Advisor: Prof. Wei Shen

  • Conducted a survey of commonly used visual recognition algorithm frameworks and carried out multi-dimensional, multi-scale experiments on these frameworks to assess the strengths and weaknesses of popular deep learning algorithms in applied downstream tasks.

  • Surveyed common visual recognition algorithms, including DETR, convolutional networks, and mask recognition networks. We analyzed these algorithms for a variety of visual recognition tasks, such as small and large object recognition, multi-task recognition, and visual object detection, culminating in a comprehensive review and comparison.

  • Compared mainstream recognition algorithms comprehensively, establishing a foundation for specific downstream task development and earning excellent evaluations in the internal review stage.

IoT

• Design and Development of Intelligent IoT Ultra-Low Power Nodes | RA        Sep 2022 - June 2023

  • Advisor: Prof. Xiaohua Tian

  • Proposed a new ultra-low power software radio system capable of minimal functionalities and OTA (Over-The-Air) remote communication.

  • Developed a compact, ultra-low power software radio based on the LoRa protocol to address size and convenience issues, modifying the node for long-distance OTA update capabilities.

  • Addressed size and portability issues in software radios with our proposed solution, significantly enhancing expandability and extending ultra-low power features to reduce power consumption to the milliwatt level.