基于HOG-CSLBP及YOLOv2的行人检测

Pedestrian detection based on HOG-CSLBP and YOLOv2

使用传统的YOLOv2网络训练出来的行人检测模型在背景简单以及行人遮掩不严重的情况下,检测效果良好,但是当背景复杂以及行人遮掩严重的时候,检测效果较差。针对此问题,在YOLOv2网络中添加HOG-CSLBP特征提取层,根据维度聚类方法对INRIA数据集目标聚类分析的结果调整YOLOv2网络的先验框个数与维度值。实验结果表明,在误检率为0.1时该算法的漏检率为9.13%,与传统的YOLOv2网络相比漏检率降低了5.27%,说明此方法有效可行。

The pedestrian detection model trained using the traditional YOLOv2 network has good detection effects in the case of a simple background and the pedestrians are not severely blocked.However,when the background is complex and the pedestrians are severely blocked,the detection effect of the model is poor.To solve this problem,the HOG-CSLBP feature extraction layer was added into YOLOv2 network,and the anchor boxes number and dimension value of YOLOv2 network were adjusted according to the results of clustering analysis of INRIA data set target.Experimental results show that the leakage rate of the proposed method is 9.13% when the false detection rate is 0.1.Compared with the traditional YOLOv2 network,the leakage rate is reduced by 5.27%,which indicates that the method is effective and feasible.