【作者】 王晨；

【导师】 曾辽原；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 随着无人机技术的快速发展,对无人机的检测与跟踪需求日益增长。其检测与跟踪面临无人机目标较小、复杂背景干扰及遮挡等挑战。为解决这些问题,本文旨在研究复杂背景下的低空无人机检测与跟踪算法,提高检测和跟踪算法的性能。具体研究内容如下:首先,本文提出了一种轻量高效的无人机检测网络NFE-YOLO。该网络设计了一种高效的正交通道注意力模块EOrthoNet,并在颈部结构中引入P卷积和C3Faster模块,以增强小目标的检测能力,提升特征提取精度,同时减少模型尺寸。此外,本文构建了一个多场景的可见光低空无人机数据集,共包含10099张精确标注的图像,以增强模型的泛化能力。实验结果表明,本文提出的NFE-YOLO在自建数据集上的m AP50值达到0.987,相较于YOLOv8n提高了2.3%,且模型大小仅为1.95MB,充分验证了方法的有效性。其次,针对单无人机目标跟踪,本文提出的改进CSRT跟踪算法融合了HOG特征、CN特征及改进的Canny边缘特征,以提升特征表达能力,并通过目标状态信息智能判断遮挡情况,结合NFE-YOLO目标检测算法实现目标重定位,从而提高跟踪的持续性和稳定性。实验结果表明,该算法在遮挡环境下表现优异,跟踪精度和成功率均突破64%,显著优于传统相关滤波算法,且在所有测试属性上均保持领先。尽管在快速运动场景中的表现逊于Siam RPN++和KPsiamfc,但整体性能良好,尤其在遮挡情况下展现出显著优势。最后,针对多无人机目标跟踪,本文提出的改进Byte Track算法在目标关联准确性方面具有明显优势。在空中场景下其MOTA指标较SORT、Deep SORT和Bo T-SORT有很大的提升,较原始Byte Track也提升2.5%。同时,IDF1指标较其他算法均有明显提高,并且ID跳变次数最少,仅发生7次,有效降低了身份切换错误。本文的研究方法为低空无人机的检测与跟踪提供了新的思路,相对于其他的方法都有所提升。更多还原

【Abstract】 With the rapid advancement of unmanned aerial vehicle(UAV)technology,the demand for UAV detection and tracking is increasing daily.However,these processes face significant challenges due to factors such as the small size of UAV targets,complex background interference,and occlusion.To address these issues,this paper aims to investigate detection and tracking algorithms specifically designed for low-altitude unmanned aerial vehicles operating in complex environments,with a focus on enhancing their performance.The specific research contributions are outlined as follows:First,we propose a lightweight and efficient UAV detection network named NFE-YOLO.This network incorporates an effective orthogonal channel attention module called EOrthoNet and integrates P convolution along with the C3Faster module within its neck structure.These innovations aim to enhance the detection capabilities for small targets while simultaneously improving feature extraction accuracy and reducing model size.Furthermore,we have constructed a multi-scene visible light dataset for low-altitude unmanned aerial vehicles that comprises 10,099 precisely labeled images.This dataset is intended to bolster the generalization ability of our proposed model.Experimental results demonstrate that the m AP50 value achieved by NFE-YOLO on our self-constructed dataset reaches 0.987—an improvement of 2.3%over YOLOv8n—while maintaining a compact model size of only 1.95MB.These findings substantiate the effectiveness of our proposed methodology.Secondly,for the tracking of single unmanned aerial vehicle(UAV)targets,the enhanced CSRT tracking algorithm proposed in this paper integrates Histogram of Oriented Gradients(HOG)features,Color Names(CN)features,and improved Canny edge features to augment feature representation capabilities.It intelligently assesses occlusion situations based on target state information and incorporates the NFE-YOLO target detection algorithm to facilitate target repositioning.This approach significantly enhances both the continuity and stability of tracking.Experimental results demonstrate that this algorithm performs exceptionally well in occluded environments,with both tracking accuracy and success rate exceeding 64%,which is markedly superior to traditional correlation filtering algorithms.Furthermore,it maintains a leading position across all tested attributes.Although its performance in fast-moving scenarios lags behind that of Siam RPN++and KPsiamfc,its overall performance remains commendable,particularly exhibiting notable advantages in occluded conditions.Finally,regarding multi-UAV target tracking,the improved Byte Track algorithm presented in this paper shows significant advantages in terms of target association accuracy.In aerial scenarios,its MOTA metric demonstrates substantial improvement compared to SORT,Deep SORT,and Bo T-SORT;it also exhibits an increase of 2.5%relative to the original Byte Track implementation.Additionally,the IDF1 index has been considerably enhanced when compared with other algorithms while minimizing identity transitions—occurring only seven times—effectively reducing identity switching errors.The research methodology outlined herein offers novel insights into the detection and tracking of low-altitude unmanned aerial vehicles and represents an advancement over existing methods.更多还原