Receptive field enrichment network for pedestrian detection

Abstract

The current advanced pedestrian detection methods adopt feature maps with different resolutions to cover multiscale pedestrians. Despite multi-scale feature pyramid can alleviate the problems caused by scale variation, each layer used for detection has merely a fixed receptive field, which results in the defects related to pedestrians with wide range of scale and aspect ratio variation. In this paper, we propose the Receptive Field Enrichment Network (RFENet), an endto- end framework for fast and accurate pedestrian detection. Two blocks are introduced in this framework, a receptive field enrichment module (RFEM) and a hierarchy aggregation module (HAM). The former is designed to diversify receptive fields of features, so as to better adapt to pedestrians with different scales and aspect ratios. The latter is further applied to enhance the entire feature hierarchy by merging spatial information and high-level semantics from different layers simultaneously. To evaluate the effectiveness of our method, extensive experiments are conducted on CityPersons and Caltech datasets. The results show that our proposed RFENet achieves comparable performance with state-of-the-art methods.