改进型SSD道路行人目标检测算法

Modified SSD road pedestrian target detection algorithm

针对道路目标检测中行人目标检测效果不佳的问题,提出了一种改进型SSD行人目标检测算法。首先,采用具有噪声的基于密度的聚类方法(DBSCAN)结合K-means算法选取适当规格的Anchor Box,使用DBSCAN剔除样本干扰点后利用K-means确定聚类中心,根据重叠度选择适当规格的Anchor Box;然后,对SSD算法的各个特征图进行尺度不变的卷积操作构建语义信息增强的特征图,并将原始特征图与增强特征图按照Concat方式特征融合,生成SSD算法的改进特征金字塔网络;最后,充分考虑正负样本不均衡的情况,选择Focal Loss函数,并结合Locatization Loss函数修正损失函数。实验结果表明,改进型SSD算法可以提高道路行人目标检测的精度和速度,并且在客观评价上取得了良好的效果。该算法在KITTI测试集上的行人目标检测平均精度为91.17%,检测速率为41.93 fps。

For poor road pedestrian detection effect in target detection,in this paper a modified SSD pedestrian target detection algorithm is proposed.First of all,the clustering algorithm DBSCAN+K-means is used to select the appropriate anchor box,DBSCAN is used to eliminate interference point and then according to IOU K-means is used to determine the clustering center;then the characteristics of figure with more semantic information is built by the scale invariant convolution operation for the SSD algorithm of several characteristic figure.And the original figure and enhancement characteristics figure is fused according to the way of Concat fusion to generate the SSD algorithm improved network characteristics of pyramid.Finally,give full consideration to the positive and negative samples imbalanced situation,Focal Loss function is chosen and combined with Locatization Loss to correct loss function.Experimental results show that the proposed improved SSD algorithm can improve the precision and speed of the road the pedestrian detection,and good results have been achieved on the objective evaluation.The algorithm in KITTI test sets the pedestrians on the detection average accuracy is 91.17%,detection rate is 41.93 frames per second.