Pedestrian Detection Using Regional Proposal Network with Feature Fusion

Pedestrian detection, which has broad application prospects in video security, robotics and self-driving vehicles etc., is one of the most important research fields in computer vision. Recently, deep learning methods, e.g., Region Proposal Network (RPN), have achieved major performance improvements in pedestrian detection. In order to further utilize the deep pedestrian features of RPN, this paper proposes a novel regional proposal network model based on feature fusion (RPN_FeaFus) for pedestrian detection. RPN_FeaFus adopts an asymmetric dual-path deep model, constructed by VGGNet and ZFNet, to extract pedestrian features in different levels, which are further combined through PCA dimension reduction and feature stacking to provide more discriminant representation. Then, the low-dimensional fusion features are adopted to detect the region proposals and train the classifier. Experimental results on three widely used pedestrian detection databases, i.e, Caltech database, Daimler database and TUD database, illuminate that RPN_FeaFus gains obvious performance improvements over its baseline RPN_BF, which is also competitive with the state-of-the-art methods.