基于边缘智能的工程约束下实时图像拼接技术实践与思考

摘 要：本文以本科生科研项目“基于边缘智能的城市安防实时街景图像拼接技术研究”为背景，梳理项目实施过程中遇到的工程问题。文章不侧重算法性能分析，而围绕 CUDA 依赖冲突、Python 版本兼容、开源库适配和设备限制等影响开发进度的关键问题，探讨其产生原因及解决方法。结合 UDIS++ 双阶段模型在 CPU 边缘设备上的优化实践，进一步总结反向问题求解、隐性假设分析等工程方法，并反思传统工科教学中重理论、轻实践的问题。本文还提出了以问题拆解、试错验证、复盘巩固为核心的三阶实践教学方法，并尝试构建涵盖环境工程能力、版本管理能力、异常处理能力的工科基础能力体系，旨在为信息时代新工科人才的工程素养培养提供一份可参考的实践样本。

关键词：边缘智能；图像拼接；UDIS++模型；工程实践；工科教育改革

正文 

可下载并阅读全文PDF，请按照本文版权许可使用。
Download the full text PDF for viewing and using it according to the license of this paper.

Practice and Reflections on Real-time Image Stitching Technology Under Engineering Constraints Based on Edge Intelligence

Abstract: In this paper, the main problems in the implementation of the project are summarized and analyzed based on the undergraduate research project " Research on real-time street scene image mosaic technology of urban security based on edge intelligence. " This paper does not focus on the performance of the algorithm, but focuses on the engineering problems that affect the progress of system development. Combined with the optimization practice of UDIS + + two-stage model on CPU edge devices, this paper summarizes engineering analysis methods such as reverse problem solving and implicit hypothesis mining. At the same time, the problems of emphasizing theory, neglecting practice, emphasizing results, neglecting process and ignoring practical constraints in traditional engineering teaching are considered. This paper also proposes a three-stage practical teaching method with problem disassembly, trial and error verification, and re-disk consolidation as the core, and tries to construct an engineering basic ability system covering environmental engineering ability, version management ability, and exception handling ability. The aim is to provide a practical sample for the engineering literacy training of new engineering talents in the information age.

Keywords: Edge Intelligence; Image Stitching; UDIS++ Model; Engineering Practice; Engineering Education Reform

参考文献

1. 施巍松，张星洲，王一帆等。边缘计算：现状与展望 [J]. 计算机研究与发展，2019, 56 (1): 69-89.
2. 陈鹏，胡斌，王力。边缘智能在城市安防视频监控中的应用研究 [J]. 计算机工程与应用，2022, 58 (12): 203-211.
3. Nie L, Lin C, Liao K, et al. UDIS: Unsupervised Deep Image Stitching for Untrimmed Images [C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 4862-4871.
4. 周济，赵恒，王永刚。深度学习图像拼接技术研究进展 [J]. 自动化学报，2023, 49 (5): 1015-1036.
5. 李泽宇，吴世忠，郭涛。开源软件依赖冲突问题研究综述 [J]. 软件学报，2021, 32 (8): 2428-2452.
6. Liao K, Lin C, Nie L, et al. Deep Image Stitching With Global and Local Mesh Alignment [J]. IEEE Transactions on Image Processing, 2021, 30: 7052-7065.
7. 韩银和，李华伟，李晓维。边缘端深度学习模型轻量化技术研究进展 [J]. 计算机学报，2022, 45 (4): 779-808.
8. 陈亮，张旭东，马洪兵。面向 CPU 端的实时深度学习推理优化技术 [J]. 清华大学学报 (自然科学版), 2023, 63 (2): 217-226.
9. 施炜，陈劲。约束驱动创新：资源限制下的企业创新逻辑与路径 [J]. 管理世界，2020, (11): 122-141.
10. 林健，彭林法，易树平。新工科人才培养的核心能力与实践教学体系构建 [J]. 高等工程教育研究，2021, (1): 17-24.
11. 顾佩华，胡文龙，林鹏。从 CDIO 到 OBE: 工程教育改革的中国路径 [J]. 高等工程教育研究，2019, (2): 1-9.
12. 吴爱华，杨秋波，郝杰。新工科建设的核心内涵、特征与推进路径 [J]. 高等工程教育研究，2020, (1): 18-25.
13. 张炜，王沛，刘拓。问题导向式学习在工科实践教学中的应用研究 [J]. 中国高等教育，2022, (12): 45-47.
14. 李培根，许晓东，陈国松。工科实践教学改革：从验证性实验到探究性实践 [J]. 高等工程教育研究，2023, (2): 1-7.
15. 教育部。教育部关于加快建设高水平本科教育 全面提高人才培养能力的意见 [EB/OL]. http://www.moe.gov.cn/srcsite/A08/s7056/201810/t20181017_351887.html, 2018-10-08.