基于YOLOV3模型的甘蔗丛环境下行人检测方法

Detection Method of Pedestrians in Sugarcane Bush Environment Based on YOLOV3 Model

针对在“一垄双沟”甘蔗丛高秆作物环境下行人目标检测的重要性及难度问题,提出一种改进YOLOV3算法的甘蔗丛中高秆作物遮挡的行人目标检测算法(YOLOV3-my-prune)。为获得更好的行人特征表达,使用Imgaug库对自制的数据集增强,并结合数据集中行人尺寸特点,使用K-means聚类分析方法,为神经网络重新聚类目标锚箱。对YOLOV3网络改进,设计了在网络全连接层中引入空间金字塔池化模块,并增加第四尺度预测特征图,同时增强网络多尺度特征融合能力及小尺度特征提取能力。完成改进的模型基础训练后,采用通道和层剪枝混合剪枝方法轻量化模型,并在数据集上进行多尺度训练并测试,结果表明:此方法准确率达80.1%,检测速度为30fps,均有所提升。

Aiming at the importance and difficulty of pedestrian detection in the sugarcane cluster environment of one ridge and double ditch,this paper proposes an improved YOLOV3 algorithm to detect pedestrians in different depths in sugarcane clusters(YOLOV3-my-prune).In order to obtain a better expression of pedestrian characteristics,imgaug library is used to enhance the data set of self-made data set,combined with the characteristics of pedestrian size in the data set,and K-means clustering analysis method is used to re-cluster the target anchor boxes for the network.For YOLOV3 network improvement,the design introduces a spatial pyramid pooling module in the network fully connected layer,adds a fourth-scale prediction feature map,and enhances the network's multi-scale feature fusion capability and small-scale feature extraction capability.After completing the improved basic training of the model,the hybrid pruning method of channel and layer pruning is adopted to lighten the model.Multi-scale training and testing on the data set show that the method in this paper has improved accuracy and detection speed.