基于超分辨率特征的小尺度行人检测网络研究

Small-scale pedestrian detection network based on super-resolution feature

为提高路侧交通场景下对小尺度行人目标的检测性能,研究了一种利用超分辨率(SR)特征的小尺度行人检测网络。首先,基于生成对抗思想学习不同尺度行人目标间的特征残差,从而将小尺度行人特征重建为与大尺度行人特征相似的超分辨率特征;并以两步检测网络Faster R-CNN为基本框架,结合特征表征性强且参数少的Inception_v2卷积层结构,采用适应小尺度行人目标的锚框参数和区域特征聚集策略ROI Align;基于超分辨率化的浅层特征实现了对小尺度行人的快速有效检测。最后,利用在路侧视角下采集的行人数据集进行了网络训练和性能测试。实验结果表明:与基准Faster R-CNN相比,提出网络对小尺度行人目标的检测准确率提升了14.7%,召回率提升了24.9%,检测速度提升至10 fps以上。

In order to improve performance of detection on small-scale pedestrian in roadside traffic scene,a small-scale pedestrian detection network using super resolution(SR)characteristics is studied.Firstly,based on generative antagonistic thought,feature residual between pedestrian target with different scale is learned,so as to rebuild small scale pedestrian feature to super resolution feature similar to large scale pedestrian feature,and two-step detection network Faster R-CNN is used as basic framework,combines Inception_v2 structure,with strong characteristic characterization and less parameters,anchor box parameters which fits small scale pedestrian target and regional feature aggregation strategy ROI Align are used.Fast and effective detection on small scale pedestrian is realized based on super-resolution characteristics of shallow.Finally,pedestrian dataset captured from the roadside view is used for network training and characteristics test.Experimental result demonstrate that compared with the baseline Faster R-CNN,detection accuracy and recall rate of the proposed network on small-scale pedestrians are increased by 14.7%and 24.9%,respectively,and the detection speed is improved to over 10 fps.