【作者】 王健；

【导师】 颜露新；

【作者基本信息】 华中科技大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 基于嵌入式平台的机载地面目标检测在电力巡检、交通监控和战场侦察等军民领域有着重要的应用。随着以深度神经网络为代表的人工智能技术快速发展,机载地面目标检测深度学习方法取得了优于传统方法的性能。但是,深度神经网络参数量大、计算密度高,而机载嵌入式平台重量、体积、功耗等资源有限,难以满足目标检测高时效性要求。为此,本文研究了目标检测深度学习模型的轻量化与优化部署技术,设计开发了基于Hi3559A的机载地面目标检测实时系统软件,具体内容如下:首先,研究了目标检测轻量化算法及部署优化技术。针对YOLO V3网络参数量大、运算量高的特点,本文优化特征提取子网络与特征融合子网络结构设计,通过深度可分离卷积技术,实现了目标检测网络结构轻量化,相比于YOLO V3模型参数量减少了83%,计算量减少了64.04%,精度仅损失3.3%;针对网络模型浮点计算复杂度高、资源耗费多的问题,结合Hi3559A平台计算特点,通过指数型模型量化与算子合并技术优化嵌入式端网络模型,降低前向推理存储与访存资源需求,轻量化网络模型部署在基于Hi3559A的嵌入式平台,在精度损失小于2%的条件下,模型大小压缩了72%,卷积运算访存量减少了75%。其次,设计开发了嵌入式实时系统软件,完成算法的嵌入式部署。针对轻量化网络模型部署后推理时间长而难以满足实时性要求的问题,通过多线程并行结构设计提升系统处理性能,启用算法处理线程并行后,达到了2倍计算加速比;进一步,研究线程实时调度策略与内存池技术,优化软件运行效率与稳定性,完成了机载地面目标检测算法的全流程部署与实现,在640×640序列图像输入条件下,地面目标检测精确度83.3%,处理帧频达到30Hz,Hi3559A处理器平台功耗4.87W,满足系统应用要求。更多还原

【Abstract】 Ground object detection based on airborne embedded platform plays an important role in many applications including power inspection,traffic surveillance,and battlefield reconnaissance.With the rapid development of artificial intelligence technology,represented by the deep neural network,the deep learning methods outperform traditional ones in ground object detection.However,with the limitations in weight,volume,and power consumption,the airborne embedded platform can hardly satisfy real-time requirements when adopting deep models with large quantities of parameters and highly computational density.To solve the above problems,this thesis implements the lightweight and optimal deployment of the deep neural network model of object detection,designs and develops Hi3559A platform-based real-time airborne ground object detection system software.The main work is as follows.Firstly,the object detection model lightweight method and deployment optimization technology are studied.In order to overcome the problem that the YOLOV3 network has huge parameters quantity and high calculation amount,this thesis lightweights the structure of the feature extraction sub-network and feature fusion sub-network,and realizes a lightweight object detection model by introducing depthwise separable convolution.Compared with the original YOLOV3 model,the lightweight model achieves an 83%reduction in parameters,a 64.04% reduction in calculation,with only 3.3% accuracy decrease.Aiming at the high complexity and high resource consumption of floating-point calculation of network models,this paper combines the calculation characteristics of the Hi3559A platform and reduces the neural network storage and memory access requirements by optimizing the embedded network model with model quantization and operator merging technology.Compared with the lightweight network model,the embedded network model deployed on the Hi3559A platform receives a 72% reduction in the model volume and a 75%reduction in convolution operation memory access,with less than 2% accuracy loss.Secondly,design and develop the embedded real-time system software and complete the deployment of the object detection algorithm.Due to the neural network’s time-consuming calculation process,it is difficult to meet the real-time requirements of the system.To address this problem,this thesis designs a multi-thread parallel structure and achieves a 2x speedup when enabling the algorithm processing threads parallelizing.Furthermore,this paper carries out research on the thread real-time scheduling strategy and memory pool technology,optimizes the software operation efficiency and stability,and completes the whole process of deployment and implementation of the airborne ground object detection algorithm.The experimental results show that under the condition of 640×640 sequence image input,the system achieves an 83.3% accuracy,and a 30 Hz processing frame rate.The Hi3559A power consumption is only 4.87 W.In summary,the system performance meets the system application requirements.更多还原