改进YOLOv3的行人车辆目标检测算法

针对YOLOv3(you only look once version 3)对中小目标检测效果不理想的问题,提出改进算法DX-YOLO(densely ResneXt with YOLOv3).首先对YOLOv3的特征提取网络Darknet-53进行改进,使用ResneXt残差模块替换原有残差模块,优化了卷积网络结构;受DenseNet的启发,在Darknet-53中引入密集连接,实现了特征重用,提高了提取特征的效率;根据数据集的特点,利用K-means算法对数据集进行维度聚类,获得合适的预选框.在行人车辆数据集Udacity上进行实验,结果表明:DX-YOLO算法与YOLOv3相比,平均准确率(mean average precision,mAP)提升了3.42％;特别地,在中等目标和小目标上的平均精度(average precision,AP)分别提升了2.74％和5.98％.

An Improved Algorithm of Pedestrian and Vehicle Detection Based on YOLOv3

Considering that YOLOv3 is not ideal for small and medium targets detection, an improved algorithm DX-YOLO is proposed. Firstly, the feature extraction network of YOLOv3 called Darknet-53 is improved, and the original residual module is replaced by ResneXt residual module, which optimizes the structure of convolution network. Inspired by Densenet, dense connection is introduced into Darknet-53 to realize feature reuse and improve the efficiency of feature extraction. According to the characteristics of data set, K-means algorithm is used to cluster the dimensions of data set to get the appropriate anchor box. Experiments on Udacity data set show that compared with YOLOv3, DX-YOLO algorithm improves the mAP by 3.42%; especially, the AP on medium and small targets increases by 2.74% and 5.98% respectively.