【作者】 王旭；

【导师】 徐勇；

【作者基本信息】 哈尔滨工业大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 随着深度学习的飞速发展,目标检测技术已经逐渐取代人工被应用在城市安防、工业制造等多个领域,人们希望目标检测技术可以越来越多地实现落地应用,从而给生活带来更多的便利。然而边缘计算设备由于性能有限,在运行深度学习目标检测模型时,通常会有较高的延迟从而无法实现实时的目标检测。针对以上问题,本文提出基于边缘计算平台的实时目标检测模型,以及面向提升推理速度的优化流程,最终设计并实现了实时目标检测系统。本文基于YOLO系列模型,设计出了高精度、低参数量的轻量化目标检测模型FYOLO(Faster YOLO)。由于模型使用轻量级网络Shuffle Netv2作为FYOLO的主干网络,使得模型的参数量得以大幅减少。在特征融合方面,引入高效的加权特征图动态融合模块,可以更好地融合来自多个不同输入的特征图;在损失融合方面,引入轻量级的自适应动态Loss权重方法,可以使得各个子任务都得到充分的学习。实验结果显示,FYOLO在COCO数据集上精度m AP达到了48.9%,而参数量只有2.7M,在边缘计算平台树莓派4B和Jetson TX2上,FYOLO在精度、速度以及参数量上做到更好的权衡。针对FYOLO模型,本文基于边缘计算平台Jetson TX2,提出了一套面向提升推理速度的目标检测流程P-FYOLO(Pipeline of FYOLO),该流程进一步提升了FYOLO在边缘计算平台上的模型推理速度。P-FYOLO使用Tensor RT推理框架,使得FYOLO模型得以实现实时检测;使用GStreamer视频图像编解码以及多线程编程等技术,使得FYOLO模型在保持识别精度不变的前提下,速度FPS额外提升了18.5%。实验结果显示,经过P-FYOLO优化推理速度后的模型在Jetson TX2上达到了45FPS左右,满足实时运行的要求。为了更好地与实际应用接轨并进行展示,本文基于P-FYOLO开发了一套实时目标检测系统,可以用于监控视角下的车流量监测以及驾驶员视角下的行人和车辆检测。同时在应用层面对模型进行拓展,设计了基于SORT追踪的区域计数法,使得车辆的检测率得以大大提升。搭配轻量级的服务器和Web界面,进一步完善了整个系统。最终测试结果表明,系统在边缘计算平台上,满足了实时、准确、实用的要求,向系统实际应用部署迈出了一大步。更多还原

【Abstract】 With the rapid development of deep learning,object detection technology has gradually replaced manual applications in many fields,such as urban security and industrial manufacturing.People hope that object detection technology can be implemented frequently,so as to bring more convenience to life.However,due to the limited performance of edge computing devices,when running the deep learning object detection model,there is usually a high delay so that real-time object detection cannot be achieved.In response to the above problems,this dissertation proposes a real-time object detection model based on edge computing platform,and an optimization process for improving inference speed.Finally,this dissertation designs and implements a real-time object detection system.Based on the series of YOLO models,this dissertation proposes a high-precision,low-parameter and lightweight object detection model FYOLO(Faster YOLO).Because the model takes the lightweight network Shuffle Netv2 as the backbone network of FYOLO,the parameters of the model can be greatly reduced.In feature fusion,a weighted feature map dynamic fusion module is proposed,which can better fuse feature maps from multiple different inputs.In order to achieve loss fusion,an adaptive dynamic loss weight method is proposed,which can make every task be fully learned.The experiments show that FYOLO achieves 48.9% accuracy on the COCO dataset,while the parameters are only 2.7M.On the edge computing platforms Raspberry Pi 4B and Jetson TX2,FYOLO has achieved better trade-off in accuracy,speed and parameters.For the FYOLO model,this dissertation proposes the object detection process PFYOLO(Pipeline of FYOLO)for improving the inference speed,based on the edge computing platform Jetson TX2.P-FYOLO uses the Tensor RT inference framework,enabling the FYOLO model to achieve real-time detection.By GStreamer video codec and multithreaded programming,the P-FYOLO make the speed(FPS)of FYOLO model increased by 18.5%,under the condition that maintain the same recognition accuracy.The experiments show that the inference speed of P-FYOLO optimized model reaches 45 FPS on Jetson TX2,which meets the requirements of real-time operation.For potential practical applications and demonstration,this dissertation proposes a real-time object detection system based on P-FYOLO,which can be used for traffic flow monitoring from the monitoring perspective,and pedestrian and vehicle detection from the driver’s perspective.In addition,the model is extended at the applications by designing an area counting method based on SORT tracking,which greatly improves the detection rate of vehicles.With the light server and web interface,the system is further improved.The final tests show that the system meets the requirements of real-time,accuracy and practicality on the edge computing platforms,and has taken a big step towards actual deployment of the system.更多还原