基于改进的YOLO与GMM相结合的地铁行人检测算法

Subway Pedestrian Detection Algorithm Based on Improved YOLO and GMM

针对目前已有的行人检测算法存在的误检率高、实时性差等问题,首先基于Yolov3-tiny网络模型进行算法改进,提出一种Ped-tiny网络模型。通过采用深度可分离卷积层替代部分原始网络中的传统卷积层来适当加深特征提取网络,同时增加一尺度的预测层,保证各个尺度的行人目标被精准检测到;其次结合GMM(混合高斯模型)的运动目标检测算法,该算法能有效利用目标运动时所产生的运动信息对行人目标进行检测、定位;最后将两算法的目标框进行对比,并对目标框进行修正。实验结果表明,在应对不同地铁场景、不同行人姿态和不同遮挡等情况时,文中方法具有更低的误检率,更高的检测精度并能满足检测的实时性要求。

Aiming at the existing pedestrian detection algorithms,such as high false detection rate and poor real-time performance,this paper improves the algorithm based on the Yolov3-tiny network model and proposes a Ped-tiny network model.The feature extraction network is appropriately deepened by using a deep separable convolution layer instead of the traditional convolution layer in part of the original network,and a prediction layer of one scale is added to ensure that pedestrian targets at all scales are accurately detected.Secondly,combined with the GMM(Hybrid Gaussian Model)moving target detection algorithm,the algorithm can effectively use the motion information generated when the target moves to detect and locate the pedestrian target.Finally,the target frames of the two algorithms are compared and the target frames are amended.