基于YOLOv2的行人检测方法研究

Research on Methods of Pedestrian Detection Based on YOLOv2

针对传统基于人工提取行人特征鲁棒性差,漏检率高的问题,借鉴目标检测的研究成果YOLOv2算法,提出基于YOLOv2网络的行人检测方法。该方法在YOLOv2网络之前加入底层特征提取层,选择LBP纹理特征作为底层特征提取层算子进行预处理,将行人背景差异转化为纹理差异,突出行人特征,然后根据行人在图片中呈现高宽比相对固定,对数据集目标框聚类分析得出最优anchor个数及维度,微调网络参数,训练得到最优模型。在INRIA行人数据集上进行试验,结果表明,该模型在行人检测中漏检率明显优于传统HOG+SVM、Faster-RCNN以及直接应用YOLOv2的方法,在INRIA数据集上误检率为10-1时,漏检率仅为9.26%。

Aiming at the problem of poor robustness and high missed detection rate based on the traditional approaches of pedestrian feature extraction,a pedestrian detection method based on the YOLOv2 network is proposed. It adds underlying feature extraction layer to YOLOv2 network before the first layer. The LBP texture feature is selected as the underlying feature extraction layer operator for preprocessing thus changing pedestrian background differences into texture differences,highlighting pedestrian characteristics. According to relatively fixed aspect ratio of the pedestrians in the picture,the optimal model in terms of numbers and dimensions of anchor will be obtained by clustering analysis of target box in dataset and fine-tuning network parameters. Experimental approaches on the INRIA pedestrian dataset have been conducted. The results demonstrate that the model in pedestrian detection performs is significantly better than the traditional HOG + SVM,Faster-RCNN and the direct application of YOLOv2 method in terms of missed detection rate. For example,when false detection rate is 0. 1,the missed detection rate is just 9. 26%.