基于全景视觉图像的实时目标检测方法  

Real-time object detection method based on panoramic image

作　　者：黄天果 何嘉[1] 沈庆阳 HUANG Tian-guo;HE Jia;SHEN Qing-yang(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《计算机工程与设计》2022年第7期2055-2061,共7页Computer Engineering and Design

基　　金：四川省科技创新苗子工程基金项目(2019Z118)。

摘　　要：针对高分辨率输入下全景视觉图像目标检测速度难以达到实时的问题,提出一种以YOLOv3为基础的轻量化网络模型,该模型在牺牲部分检测精度的条件下实现了较好的检测速度提升。为解决全景图像训练数据集样本数量不足的问题,提出一种适用于全景图像的数据增强方法来扩充训练数据集。实验结果表明,利用YOLOv3验证扩充后的数据集以交并比(IoU)为0.5的条件下评估平均精度,与原始数据集相比检测精确度提高4.75%,改进的轻量化模型与YOLOv3相比,网络参数量减少65.08%,检测速度提升31.81%并达成实时检测。To achieve real-time object detection on high-resolution panoramic images,with a minor sacrifice of accuracy,a lightweight model based on YOLOv3 was proposed for major speed improvement.To solve the lack of samples in panoramic image datasets,a data augmentation method was proposed for expanding dataset.According to experiment results,the mean average precision increases 4.75%using YOLOv3 on expanded dataset compared to original dataset,while the IoU is 0.5.65.08%parameters are reduced and the detection speed is 31.81%faster using improved lightweight YOLOv3 compared with the original YOLOv3,achieving real-time detection.

关 键 词：全景视觉图像 实时目标检测 轻量化模型 YOLOv3 数据增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]