Abstract: In complex scene, there are a lot of pedestrians who are not properly detected by the region proposal network model,which is trained with Caltech dataset. Firstly, the feature extracted by VGG model in region proposal network can’t represent pedestrian information well. Secondly, hand-crafted sizes of the anchor box don’t use the annotation information of pedestrians. In this paper, we proposed an improved pedestrian detection method of regional proposal networks to address the problems. We extracted the feature of the image by using the ResNet which have a better performance on the classification task. Then use the detection sub-network sliding on the feature map to predict whether multiple anchor boxes are pedestrian and adjust their position and size, especially the anchor boxes’ size are calculated by KMeans algorithm. The experimental show that, compared with the traditional VJ and HOG methods, the algorithm’s missed detection rate are 36.23% and 27.09% lower on Caltech dataset. Moreover, the algorithm’s missed detection rate are also 6.78%, 3.73%, 1.53% lower than PedFaster RCNN, MRFC+Semantic, and UDN+ based on the deep learning method. The results show that the improved regional candidate network can be used to detect pedestrians well.