基于改进YOLO算法的全景多目标实时检测  

Panoramic multi-object real-time detection based on improved YOLO algorithm

作　　者：蔡成涛[1] 吴科君 刘秋飞 程海涛 马强 CAI Cheng-tao;WU Ke-jun;LIU Qiu-fei;CHENG Hai-tao;MA Qiang(College of Automation,Harbin Engineering University,Harbin 150001,China;Harbin Jian Cheng Group Limited Company,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001 [2]哈尔滨建成集团有限公司,黑龙江哈尔滨150001

出　　处：《计算机工程与设计》2018年第10期3259-3264,3271,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61673129);黑龙江省自然科学基金项目(F201414)

摘　　要：针对折反射全景视觉系统特殊的成像原理导致的全景图像变形的特点,以YOLO算法为基础,提出一种E-D-YOLO(expand density YOLO)全景多目标实时检测方法。建立全景系统的成像参数模型,根据全景视觉系统模型进行投影点解算,得到待检目标"矮粗化"的柱面展开图,修改YOLO的网络结构中纵向网格数量,使其适应"矮粗化"的待检目标。实验结果表明,E-D-YOLO相对于YOLO算法在全景目标检测上具有更高的准确度,检测速度高达31帧每秒,保持可观检测精度的同时具有实时性,此外改变实验环境进行对比实验,对比结果表明,E-D-YOLO方法具有一定鲁棒性。Based on the YOLO algorithm,an E-D-YOLO(expand density YOLO)panoramic multi-target real-time detection method was proposed for the characteristics of the defocused panoramic vision imaging target deformation.The imaging parameter model of the panoramic system was established,and the projection point was calculated according to the panoramic vision system model.The cylindrical expansion of target was shortly coarsened,and the number of longitudinal grids in the network structure of YOLO was modified to adapt to the shortly coarsened target to be tested.Experimental results show that the E-D-YOLO has higher accuracy than the YOLO algorithm in the detection of the panoramic target.The detection speed is as high as 31 frames per second,and the observation accuracy is kept at the same time.In addition,the experimental environment is changed to show that E-D-YOLO method has a certain robustness.

关 键 词：全景图像 深度学习 神经网络 多目标 实时检测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]