摘要
近些年来,目标检测算法在移动机器人环境感知领域表现出了突出的性能。但是目标检测算法存在模型庞大和计算复杂的问题,制约了目标检测算法在移动嵌入式设备上的部署和发展。YOLO是一种单阶段的目标检测算法,具有较高的准确度和较快的运行速度。该文提出了一种基于YOLOv5s改进后适用于嵌入式设备的移动机器人目标检测算法DID-YOLO。首先,使用深度可分离卷积和倒置残差模块对YOLOv5s的backbone网络进行重构,降低模型复杂度和计算量,达到轻量化的目的;其次,利用特征层和输出层结合的知识蒸馏训练提高重构后目标检测网络的精度。在目标检测通用数据集PASCAL VOC上实验表明:DID-YOLO模型尺寸为3.63 MB,相较原网络模型尺寸减小了48.65%;经过特征层和输出层蒸馏后,DID-YOLO的mAP@0.5提升至73.83%;DID-YOLO在Jetson AGX Xavier上实现了每秒31.2帧的实时图像处理速度。提出的DID-YOLO性能显著,满足了移动机器人嵌入式平台的实时高精度检测需求。
In recent years,object detection algorithm has shown outstanding performance in the field of mobile robot environment perception.However,the problem of huge model and complex calculation of target detection algorithm restricts the deployment and development of such algorithm on mobile embedded devices.YOLO is a single-stage target detection algorithm with high accuracy and fast running speed.We propose an improved DID-YOLO mobile robot target detection algorithm based on YOLOv5s,which is suitable for embedded devices.Firstly,deep separable convolution and inverted residual modules are used to reconstruct the backbone network of YOLOv5s to reduce model complexity and computational load for lightweight.Secondly,knowledge distillation training which combines the feature layer and the output layer is used to improve the accuracy of the reconstructed target detection network.Experiments on PASCAL VOC show that the size of DID-YOLO model is 3.63 MB,which is 48.65%smaller than that of the original network model.After distillation of the feature layer and the output layer,mAP@0.5[JP2]of DID-YOLO increased to 73.83%.DID-YOLO achieved a real-[JP]time image processing speed of 31.2 frames per second on Jetson AGX Xavier.The DID-YOLO proposed has remarkable performance and can meet the requirements of real-time and high-precision detection of embedded mobile robot platforms.
作者
章佳琪
肖建
ZHANG Jia-qi;XIAO Jian(School of Electronic and Optical Engineering&School of Flexible Electronics(Future Technology),Nanjing University of Posts and Telecommunications,Nanjing 210046,China;School of Integrated Circuit Science and Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210046,China)
出处
《计算机技术与发展》
2023年第10期8-14,共7页
Computer Technology and Development
基金
国家自然科学基金项目(61974073)。
关键词
移动机器人
目标检测
嵌入式设备
轻量化
知识蒸馏
mobile robot
object detection
embedded device
lightweight
knowledge distillation
