摘要： 针对基于人工提取特征的行人检测器鲁棒性差的问题,借鉴目标检测的研究成果,提出一种行人目标实时检测方法。以YOLO网络结构为基础,结合行人在图像中呈现宽高比小的特点,聚类选取合适的候选框数量和规格,改进YOLO网络结构,调整候选框在X、Y轴方向的分布密度,形成适用于行人检测的网络结构。实验结果表明,与HOG、LatSVM-v2等行人检测方法相比,该方法降低了漏检率和误检率,提高了定位准确性,检测速度满足实时性要求。

关键词: YOLO网络, 行人检测, 深度网络, 聚类, 特征重组

Abstract: For improving the poor robustness of pedestrian detectors based on artificial extraction characteristics,learns from the advanced research achievements in the object detection field,proposes a method of real-time pedestrian detection.Being based on the YOLO network structure,combines with characteristics that pedestrians’ aspect ratio is small in images,selects appropriate numbers and scales of the initial candidate boxes by clustering,improves the YOLO network structure,adjusts density of candidate boxes distributed on the X and Y axis,forms a network structure which is suitable for pedestrian detection.Experiments in INRIA pedestrian datasets show that comparing with some pedestrian detection methods,such as HOG,LatSVM-v2 and so on,this method reduces the missed and false detection,improves the positioning accuracy and satisfies the real-time requirements.

Key words: YOLO network, pedestrian detection, deep network, clustering, features reorganization

中图分类号: 

• TP391