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.
Pedestrian detection has significant applications in the field of driver assistance, video surveillance,intelligent robot and so on. To address the problem of many false detections and missing windows produced by existing pedestrian detection algorithm when the videos are of low resolution with complicated background, this paper propose a fusion method with multi-models and intra-frame information for pedestrian detection. It combines complementary detection results generated from Fast R-CNN and Faster R-CNN which can obtain more precise detection windows at first. Then a frame-context information fusion method is proposed to further remove the false positives and false negatives resulting from single frame information.In the Caltech pedestrian detection dataset, under the condition of false positive per image (FPPI) equaling 10%, the missing rate can be reduced to 14.04% which is 2.05% lower than that (16.09%) of Faster R-CNN model. It shows that fusion of multi-models and intra-frame information can correct the results of previous pedestrian detection and improve the detection performance accordingly.
作者
王斌
刘洋
唐胜
郭俊波
Wang Bin;Liu Yang;Tang Sheng;Guo Junbo(Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190;National Computer Network Emergency Response Technical Team/Coordination Center of China, Beijing 100029;University of Chinese Academy of Sciences, Beijing 100049)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2017年第3期444-449,共6页
Journal of Computer-Aided Design & Computer Graphics
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)
关键词
行人检测
卷积神经网络
模型融合
帧间信息融合
pedestrian detection
convolutional neural networks
model fusion
intra-frame information fusion