Obstacle detection in complex farmland environment based on improved YOLOv5

In order to realize the rapid detection of obstacles in the process of autonomous navigation of agricultural robots in complex field environments, an obstacle detection method based on improved YOLOv5 in complex field environments is proposed. The farmland obstacle data set containing a total of 6 766 images of agricultural machinery, human and sheep objects are established. The optimal prior anchor box size is generated by the k-means clustering algorithm. The CBAM convolution block attention module is introduced to suppress the interference of the complex environment around the target obstacle and enhance the target saliency. A detection head is added to connect the backbone features across levels, enhance the ability to express multi-scale features, and alleviate the negative impact of the variance of the scale of the labeled objects. The Ghost convolution is used to replace the ordinary convolution in the Neck layer to reduce the model parameters and decrease the model complexity. Compared with the YOLOv5s benchmark model, the improved model has increased the detection accuracy by 2.3%, the recall rate by 3.1%, the accuracy rate by 1.9%, and has decreased the reference number by about 7%. It provides technical reference for the research of navigation and obstacle avoidance during autonomous operation of unmanned agricultural machinery.