基于改进Faster R-CNN的行人检测算法

Pedestrian Detection Algorithm Based on Improved Faster R-CNN

针对行人检测中复杂环境,提出一种改进Faster R-CNN的行人检测算法,使用深度卷积网络从图片中提取适合检测目标的特征。基于Faster R-CNN算法,以Soft-NMS算法代替传统NMS算法,加强Faster R-CNN算法对重叠区域的识别能力。同时,算法通过"Hot Anchors"代替均匀采样的锚点避免大量额外计算,提高检测效率。最后,将21分类问题的Faster R-CNN框架,修改成适用于行人检测的2分类检测框架。实验结果表明:改进Faster R-CNN的行人检测算法在VOC 2007行人数据集,检测效率和准确率分别提升33%、2.6%。

Aiming at the complex environment of pedestrian detection,an improved Faster R-CNN algorithm for pedestrian detection was proposed.The deep convolution network was used to automatically extract the features that are most suitable for detecting objects from the image.Based on the Faster R-CNN algorithm,soft NMS algorithm was used to replace the traditional NMS algorithm,by which the recognition ability of the Faster R-CNN algorithm for overlapping areas was enhanced.At the same time,"Hot Anchors"was used to replace the anchor points of uniform sampling to avoid very much extra calculation and improve the detection efficiency.Finally,the Faster R-CNN framework for 21-classification problem was modified to a 2-classification framework for pedestrian detection.The improved pedestrian detection algorithm was verified on VOC 2007 pedestrian data set,and its detection efficiency and accuracy were improved by 33%and 2.6%respectively.