CASIA OpenIR  > 毕业生  > 博士学位论文
面向开放环境无人平台的持续目标检测算法研究
冯航涛
2024-05-13
Pages134
Subtype博士
Abstract

在实际应用中,无人平台所处的场景通常具有开放性,与静态场景不同,开放场景中会不断出现与已知类别或者域不同的目标,这增加了无人平台长期稳定进行目标检测的难度。此外,无人平台由于自身的硬件和算力限制,以及隐私问题的考量,难以大量保存已知类别或者域的样本。因此,具有持续学习能力的目标检测技术对于无人平台的感知发育和知识增长具有非常重要的实际意义。然而,当前开放环境下无人平台的持续目标检测研究仍面临多个挑战和问题。首先,持续目标检测的评价框架和指标缺乏统一,导致难以有效比较不同算法的性能;其次,全天候场景的数据集不足,限制了算法在光线较弱或视觉不清的环境中的应用效果;特别是在数据有限的战场环境下,现有算法因依赖大量数据而难以有效快速地识别新型目标;最后,算法在实际应用中的性能评估与其在静态数据集上的测试结果存在差异,这影响了算法在动态真实世界任务中的可靠性。这些问题共同阻碍了持续目标检测技术在实际应用中的发展。因此,本研究旨在针对以上开放环境下无人平台视觉感知的问题,结合持续目标检测任务,展开面向开放环境无人平台目标检测算法的研究,具体如下:


(1)基于稳定性和可塑性度量的持续目标检测评估准则:作为无人平台视觉感知的核心环节,持续目标检测近年来受到广泛的研究和关注。然而,目前的研究通常采用多样的框架,这严重影响了算法性能的比较分析的有效性和可靠性。此外,由于缺乏专门针对持续目标检测的评价准则,多数研究仍依赖于目标检测的通用评估指标,如平均精度均值(mean Average Precision, $mAP$)。而这些指标无法全面反映模型在持续学习环境下的抗遗忘性和对新知识的适应能力。分析现有持续目标检测算法的评估方式,分别从模型的抗遗忘能力,对新知识的学习能力以及两者综合的角度对已有的评估方式进行改进,以更好地描述各类持续目标检测算法的性能和差异。在一种标准化的持续目标检测评估框架,对不同算法在相同设置下的性能进行对比分析,可较为公正地比较各算法的优劣,证明了所提出评估准则的有效性,并为算法的选择提供依据。

(2)基于无人机平台的持续目标检测数据集及模态感知门控知识蒸馏持续目标检测算法:针对无人机全天候目标检测任务,本文搭建了一种基于无人机平台的多模态数据集,结合了彩色和红外图像,包括16000对已对齐的图像,以及丰富的目标检测标签,旨在推动相关领域的研究。同时,本文根据提出的数据集,设计了一种针对多模态数据的模态感知门控知识蒸馏持续目标检测算法。该算法通过自训练方法学习多模态特征表示,通过使用与域相关的门函数来筛选与域相关的特征,以解决知识蒸馏过程中特征筛选的问题。同时,本文使用该知识蒸馏方法将知识从旧模型蒸馏到新模型中,结合旧模型生成的伪标签,提高了模型的抗遗忘能力,以降低灾难性遗忘对持续目标检测的影响。在提出的彩色-红外多模态无人机数据集上验证了算法的有效性,可以大幅度提高模型的抗遗忘能力。

(3)基于知识迁移的小样本持续目标检测算法:在面对复杂多变的开放环境时,采集大量数据是一项困难且昂贵的任务。因此,如何在少量样本条件下有效地进行持续目标检测是一个普遍存在的问题。为了解决这个问题,本文将开展基于知识迁移的小样本持续目标检测算法的研究。作者提出了两种创新性的算法模型:多类头(Multi-Class Head, MCH)模型和双路径多类头(Bi-Path Multi-Class Head, BPMCH)模型。MCH模型在检测新类时为每个新类添加一个新的分类头,对新类目标的特征进行分类检测。BPMCH模型在此基础上,为新类复制了一个新的骨干网络和特征金字塔网络(Feature Pyramid Network, FPN),并使用基类网络的参数进行初始化,以更好地传递基类知识。在训练过程中,作者先在基类上充分训练检测模型,然后在新类少量数据上微调新的分类头和骨干网络,实现增量学习。此外,本文在头部结构中设计了新旧特征交互机制,充分利用基类和新类的特征信息,进一步提高算法性能。在具有挑战性的MS COCO数据集上显著提升了检测效果,在精度上超越了当时性能最好的方法。

(4) 开放环境无人机视觉感知系统搭建及算法验证: 将理论方法和关键技术应用于实际生产和生活,进而解决真实场景中的关键问题,具有重要的意义和价值。本文以现有的持续目标检测技术为基础,结合新搭建的共轴双旋翼无人机视觉感知系统,利用目标检测算法、迫近下降算法、薄弱位置定位和6D姿态估计算法完成了典型的军事场景任务。并且结合该平台,本文也证实了所提出的持续目标检测方法在真实开放场景中的有效性,为相关领域的研究和实践奠定了坚实的基础。

Other Abstract

In practical applications, unmanned platforms typically operate in open environments, where, unlike in static scenarios, novel objects of different classes or domains continuously emerge, posing challenges to the long-term stable detection of objects. Additionally, constraints such as hardware limitations, computational power, and privacy concerns make it difficult for unmanned platforms to store a large number of samples belonging to known classes or domains. Therefore, object detection technology with continual learning capability is of great practical significance for the perception development and knowledge growth of unmanned platforms. The current research on continual object detection in open environments faces several challenges: firstly, the lack of a unified evaluation framework and metrics makes it difficult to effectively compare the performance of different algorithms; secondly, the inadequacy of datasets for all-weather scenarios restricts the application effectiveness of algorithms in low-light or visually unclear environments; furthermore, particularly in data-limited battlefield environments, existing algorithms struggle to efficiently and rapidly identify novel objects due to their reliance on extensive data; finally, the disparity between algorithm performance evaluation in real-world applications and testing on static datasets affects the reliability of algorithms in dynamic real-world tasks. These issues collectively impede the progress and widespread adoption of continual object detection technology in practical applications. Therefore, this study aims to address the aforementioned challenges in unmanned platform visual perception in open environments, by conducting research on object detection algorithms tailored for open environment unmanned platforms, specifically focusing on continual object detection tasks. The specific objectives are as follows:

(1) Evaluation Criteria for Continual Object Detection Based on Stability and Plasticity Measures: As a core component of visual perception for unmanned platforms, continual object detection has garnered significant research interest in recent years. However, current studies often employ diverse frameworks, which severely impacts the effectiveness and reliability of performance comparisons among algorithms. Additionally, due to the lack of evaluation criteria specifically designed for continual object detection, most research still relies on general evaluation metrics for object detection, such as mean Average Precision (mAP). These metrics fail to comprehensively reflect a model's ability to resist forgetting and adapt to new knowledge in a continual learning environment. Analyzing the evaluation methods of existing continual object detection algorithms, we propose improvements from the perspectives of the model's ability to resist forgetting, its ability to learn new knowledge, and a combination of both. This aims to better describe the performance and differences of various continual object detection algorithms. By establishing a standardized evaluation framework for continual object detection, we can fairly compare the performance of different algorithms under the same settings, demonstrating the effectiveness of the proposed evaluation criteria and providing a basis for algorithm selection.

(2)Continual Object Detection Dataset and Modal-aware Knowledge Distillation Algorithm on UAV Platform: This paper presents a multi-modal dataset based on unmanned aerial vehicle (UAV) platforms for all-weather object detection tasks, integrating both color and thermal images. The dataset comprises 16000 aligned image pairs accompanied by rich object detection annotations, aimed at advancing research in related fields. Additionally, leveraging the proposed dataset, this paper devises a modality-aware gate-controlled knowledge distillation algorithm for continual object detection in multi-modal data. The algorithm learns multi-modal feature representations through self-training methods, employing modality-relevant gate functions to filter domain-specific features during the knowledge distillation process. Moreover, utilizing this knowledge distillation approach, knowledge is transferred from the old model to the new model, augmented with pseudo-labels generated by the old model, thereby enhancing the model's resistance to catastrophic forgetting to mitigate its impact on continual object detection. The effectiveness of the algorithm is validated on the proposed color-thermal multi-modal UAV dataset, demonstrating significantly improved resistance to catastrophic forgetting.

(3) Few-shot Continual Object Detection Algorithm based on Knowledge Transfer: In the face of complex and volatile open environments, collecting large amounts of data is a difficult and costly task. Therefore, effectively conducting continual object detection with a small number of samples is a common problem. To address this issue, this paper explores the research of small-sample continual object detection algorithms based on knowledge transfer. The authors propose two innovative algorithm models: the Multi-Class Head (Multi-Class Head, MCH) model and the Bi-Path Multi-Class Head (Bi-Path Multi-Class Head, BPMCH) model. The MCH model adds a new classification head for each new class when detecting new classes, classifying and detecting features of new class objects. Building upon this, the BPMCH model duplicates a new backbone network and Feature Pyramid Network (Feature Pyramid Network, FPN) for new classes, initializing them with parameters from the base class network to better transfer base class knowledge. During training, the authors fully train the detection model on the base class first, then fine-tune the new classification head and backbone network on a small amount of data for new classes, achieving incremental learning. Additionally, this paper designs a mechanism for feature interaction between old and new features in the head structure, fully utilizing the feature information of base and new classes, further improving algorithm performance. Significant improvements in detection performance were achieved on the challenging MS COCO dataset, surpassing the performance of the best methods at the time.

(4) Practice of Continual Object Detection in Open Environment UAV Platform Scenarios: The application of theoretical methods and key technologies to real-world production and everyday life, addressing critical problems in practical scenarios, is of profound significance and value. This article is based on existing continual object detection technology, combined with a newly constructed coaxial dual-rotor unmanned aerial vehicle (UAV) visual perception system. Using object detection algorithms, approach descent algorithms, weak position localization, and 6D pose estimation algorithms, the system efficiently completes tasks in typical military scenarios. Furthermore, using this platform, the paper validates the effectiveness of the proposed continual object detection method in real and open environments, thereby laying a solid foundation for further research and practical applications in related fields.

Keyword持续学习 目标检测 持续目标检测 无人平台
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56498
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
冯航涛. 面向开放环境无人平台的持续目标检测算法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
明版论文-冯航涛-0528.pdf(12990KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[冯航涛]'s Articles
Baidu academic
Similar articles in Baidu academic
[冯航涛]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[冯航涛]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.