Pedestrian Detection based on Deep Fusion Network using Feature Correlation

Abstract

Since most of the pedestrian detection method focus on color images, the detection accuracy is lower when the images are captured at night or dark. In this paper, we propose a deep fusion network based pedestrian detection method. We utilize deconvolutional single shot multi-box detector (DSSD) fused at halfway stage. Also, we apply feature correlation for two image modality feature maps to produce a new feature map. For the experiment, we use KAIST dataset to train and test the proposed method. The experiment results show that the proposed method gains 22.46% lower miss rate compared to the KAIST pedestrian detection baseline. In addition, the proposed method shows at least 4.28% lower miss rate compared to the conventional halfway fusion method.