融合多模型和帧间信息的行人检测算法

发布时间：2018-05-08 00:26

本文选题：行人检测 + 卷积神经网络　；参考：《计算机辅助设计与图形学学报》2017年03期

【摘要】：行人检测在车辆辅助驾驶、视频监控、智能机器人等领域具有重要的应用价值.针对当前行人检测算法在视频分辨率低和背景复杂的情况下存在很多误检及漏检的问题,提出一种融合多模型和帧间信息的行人检测算法.首先融合Fast R-CNN和Faster R-CNN模型的互补检测结果获取精准的检测窗口;然后采用视频帧间上下文融合算法来弥补单帧图像检测算法存在的漏检和误检.实验结果表明,在Caltech行人检测数据库上,在每幅图像虚警率(FPPI)为10%的条件下,该算法丢失率仅为14.04%,比Faster R-CNN单模型丢失率(16.09%)降低2.05%;利用多模型和帧间信息融合对行人检测结果进行校正,能提高行人检测性能.
[Abstract]:Pedestrian detection has important application value in vehicle aided driving, video surveillance, intelligent robot and so on. In view of the problems of false detection and missed detection in current pedestrian detection algorithms with low video resolution and complex background, a pedestrian detection algorithm combining multi-model and inter-frame information is proposed. Firstly, the complementary detection results of Fast R-CNN and Faster R-CNN model are fused to obtain accurate detection window, and then the video frame context fusion algorithm is used to compensate for the missing and false detection of single frame image detection algorithm. The experimental results show that in the Caltech pedestrian detection database, the false alarm rate of each image is 10%. The loss rate of this algorithm is only 14.04, which is 2.05 lower than that of Faster R-CNN single model (16.09). The multi-model and inter-frame information fusion is used to correct the pedestrian detection results, and the pedestrian detection performance can be improved.
【作者单位】：中国科学院计算技术研究所智能信息处理重点实验室;中国科学院大学;国家计算机网络应急技术处理协调中心;
【基金】：国家自然科学基金(61572472,61525206) 国家“八六三”高技术研究发展计划(2014AA015202) 北京市自然科学基金(4152050)
【分类号】：TP391.41

【相似文献】