Pedestrian Detection Based On Semi-supervised Region Proposal Network

In this era where big data and artificial intelligence are prevalent,techniques of pedestrian detection are widely applied on the areas of autopilot,video surveillance and personnel retrieval.With the rapid growth of cameras,it is relatively easy to collect massive data through cameras,but manually labeling a large amount of training data for each scenario is expensive.At present,the mainstream pedestrian detection methods are based on the Convolutional Neural Networks(CNN),and the training process usually requires a great amount of labeled data.To address this problem,semi-supervised pedestrian detection studies how to train a detection model using a little labeled data and exploiting a large amount of unlabeled data.Existing semisupervised methods generally adopt the multi-stage training strategy,which requires a high time cost.Aiming at the issues of the detection models' reliance on a large number of labeled data and the high time cost of multi-stage strategy training,based on the Region Proposal Network(RPN)of Faster Region-based Convolutional Neural Networks(Faster R-CNN),we propose the Self-Enhanced Region-based Convolutional Neural Networks(SE-RCNN)model to improve semi-supervised pedestrian detection performance.The SE-RCNN model consists of two modules,namely Reliability Analysis(RA)and Self-Enhanced Detection(SED).In order to utilize unlabeled data,the SED module extracts high-confidence proposals and the RA module is responsible for filtering the low-quality proposals.The remaining high-quality proposals took as pseudo-labels are to improve the generalization performance of the SED module.The RA module is composed of a saliency detection branch and an Intersection over Union(Io U)estimation branch.The saliency detection branch realizes the perception of the pedestrian saliency information and encodes this information into a shared feature map.The Io U estimation branch evaluates the position of the human body in the proposals,excluding situations such as too large,too small,or off-center area occupied by the human body,so as to ensure the accuracy of the pseudo-labels.Therefore,a little labeled data and a large amount of pseudo-labeled data can be used for model optimization simultaneously to achieve the purpose of single-stage training detector.On a number of pedestrian detection benchmark datasets,we verify that the RA module can effectively improve the accuracy of pseudo-label data,and can also improve the detection performance of the SED module.The proposed SE-RCNN model outperforms all the comparison methods.