红外技术, 2018, 40 (6): 578, 网络出版: 2018-08-04
基于改进Fast R-CNN的红外图像行人检测研究
Research on Infrared Image Pedestrian Detection Based on Improved Fast R-CNN
快速区域卷积神经网络 红外图像 行人检测 自适应ROI 提取 加权锚点框 fast region convolution neural network infrared image pedestrian detection adaptive ROI extraction weighted anchor box
摘要
针对红外图像行人检测任务中行人细节信息少,特征提取计算量大以及易受背景影响等问题,提出了一种改进的Fast R-CNN(快速区域卷积神经网络)红外图像行人检测方法。改进主要涉及两个方面:①结合红外图像的特点提出了一种自适应ROI 提取算法,在不影响检测准确率的前提下,降低了ROI 数量,使得网络的计算量减小;②提出了一种加权锚点框的定位机制,基于3 种不同宽高比锚点框的检测置信度进行坐标加权,获得更准确的定位框。实验结果表明,本文提出的改进方法与传统的Haar+LBP+HOG+SVM 算法及Fast R-CNN 算法相比,红外图像行人检测的准确率从80.3%和91.2%提高到92.3%,检测速度从68 ms/f 和25 ms/f 提高到12 ms/f,提高了系统的性能。
Abstract
To solve the problems of infrared pedestrian detection, such as fewer details of pedestrian information, large amount of feature extraction calculation and easily affected by background, a detection method based on improved fast region convolution neural network was proposed. The improvements include two aspects: firstly, an adaptive ROI extraction algorithm was proposed based on the characteristics of infrared image. On the premise of not affecting the detection accuracy, the number of ROI was reduced and the computation of the network was reduced. Secondly, a weighted anchor box location mechanism was proposed, which was based on three kinds of detection confidence degree of anchor bounding-box with different aspect ratios, and coordinates were weighted to get a more accurate bounding-box. Experimental results showed that the accuracy of pedestrian detection in infrared images was increased from 80.3% and 91.2% to 92.3% by both improved methods proposed in this paper. Compared with the traditional Haar +LBP+HOG+SVM algorithm and Fast R-CNN algorithm, the detection speed was increased from 68 ms/f and 25ms/f to 12ms/f, the performance of the system is improved.
车凯, 向郑涛, 陈宇峰, 吕坚, 周云. 基于改进Fast R-CNN的红外图像行人检测研究[J]. 红外技术, 2018, 40(6): 578. CHE Kai, XIANG Zhengtao, CHEN Yufeng, LYU Jian, ZHOU Yun. Research on Infrared Image Pedestrian Detection Based on Improved Fast R-CNN[J]. Infrared Technology, 2018, 40(6): 578.