Long-distanceinfrared video pedestrian detection using deep learning and backgroundsubtraction

Abstract

Infrared video-based pedestrian detectionplays an important role in surveillance and automatic driving. Compared with vision cameras, infrared cameras have the characteristics of all-day and all-weather availability. The current infrared video detection methods generally follow the traditional paradigm. However, when objectsare at a long distance, it is challenging for existing methods to achieve accurate detection. Meanwhile, deep learning-based methods have been widely used in visible light object detection. In this paper, we introduce an algorithm which combines deep learning and background subtraction methods to achieve infrared pedestrian detection. Firstly, background subtraction method is performed to provide the inter-frame information for the deep learning module. Secondly, the RefineDet equipped with an attention module is used to improve the detection accuracy for small pedestrian. Moreover, we develop a novel infrared video dataset which includes long-distance pedestrian for performance evaluation. Experiments demonstrated that our method can achieve a superior performance as compared to RefinDetand SSD methods.