ADN: ATSS-based Deep Network for Pedestrian Detection on High-Resolution Images

Abstract

This paper investigates the high-resolution pedestrian detection network AND(ATSS-based Deep Network for Pedestrian Detection on High-Resolution Images) to solve the problem that high-resolution images are difficult to be processed directly as well as pedestrian occlusion and pose variations. First, it obtains a deeper backbone network by stacking residual modules to extract multi-level features, improve the extraction ability of occluded target features and avoid network degradation. Subsequently, the deformable convolution is introduced to optimize the backbone network and expand the local receptive field, thus further optimizing the detection ability of deformable targets and occluded targets. On the PANDA dataset, the AP, AP50, and AP75 of ADN are 3.5, 2.7, and 4.2, which are higher than that for the baseline respectively. Compared with other state-of-the-art methods, experiments show that ADN effectively enhances the accuracy of pedestrian detection in high-resolution Imgaes, and it outperforms existing object detection algorithms.