Zenithal People Detection Based on Improved Faster R-CNN
Lin Li, Jinquan Ma · 2018
With the development of computer vision, zenithal people detection has attracted much attention in our society. Recent algorithms have shown effective performance which are mostly based on machine learning and hand-crafted features. However, the detection of the surveillance with large and various pedestrian flow still has limited success due to the head irregularity and occlusion. To address this problem, in this paper, we propose an improved zenithal detection network based on Faster R-CNN. We extend the network to a more discriminative and effective one by adopting a “joint-loss” function and presenting a new combination of the anchors according to the characteristic of human head feature. Note that the new anchors proposed are more appropriate to the head detection task. Experimental results demonstrate that the proposed framework has substantially better accuracy of state-of-the-art methods.