基于改进YOLOv5s的田间移动障碍物检测

Detection of improved YOLOv5s based mobile obstacles in field

为实现无人农机在行驶过程中对田间移动型障碍物的实时检测,提出一种基于YOLOv5s的目标检测模型,用于检测田间行人和其他协同作业的农机设备。该目标检测模型以YOLOv5s模型为基础框架,进行了以下三点改进:第一,为了减少模型的参数量和计算复杂度,提高推理速度,将YOLOv5s网络模型中的卷积模块和C3模块替换为Ghost卷积和C3Ghost模块;第二,为了弥补模型参数量减少所造成的精度下降的损失,提升对目标的检测能力,在主干网络输出的特征层中引入CBAM注意力机制;第三,采用BiFPN特征金字塔结构,实现多尺度特征加权融合。实验结果表明,YOLOv5s模型的参数量为7.02×106,计算复杂度为15.8GB,平均检测精度为94%,生成权重文件大小为13.7MB,单幅图像的检测速度为71.43 f/s;改进后的模型参数量为4.04×106,下降了42.45%,计算复杂度缩减为8.5 GB,平均检测精度达到了93.2%,仅仅下降了0.8%,权重文件大小为8.1 MB,单幅图像的检测速度为77.52 f/s。以上数据证明,改进后的模型能够满足对田间移动型障碍物的实时检测,且更加易于部署到移动端设备。

In order to realize the real-time detection of movable obstacles in the field during the operation of unmanned agricultural machinery,an object detection model based on YOLOv5s is proposed,which is used to detect people and other agricultural machineries with collaborative operations in the farmland.In the object detection model,the YOLOv5s is used as the basic framework to conduct three improvements.In order to reduce the number of parameters and computational complexity of the model,and improve inference speed,the convolution module and C3 module in the YOLOv5s network model are replaced with Ghost convolution and C3Ghost modules.In order to compensate for the loss of accuracy caused by the reduction of model parameters and improve the detection ability of targets,CBAM attention mechanism is introduced in the feature layer output by the backbone network.The BiFPN feature pyramid structure is used to achieve multi-scale feature weighted fusion.The experimental results show that the parameter size of the YOLOv5s model is 7.02×106,the computational complexity is 15.8 GB,the average detection accuracy is 94%,the generated weight file size is 13.7 MB,and the detection speed of a single image is 71.43 f/s.The parameter size of the improved model is 4.04×106,with a decrease of 42.45%,the computational complexity is 8.5GB,the average detection accuracy is 93.2%,with a decrease of 0.8%,the generated weight file size is 8.1 MB,and the detection speed of a single image is 77.52 f/s.The above data proves that the improved model can meet the real-time detection of mobile obstacles in the field and is easier to deploy to mobile devices.