Multi-Region Attention Residual Network-Based Crop Row Detection For Agricultural Robot Field Navigation

In the context of real-world agricultural settings, robots frequently encounter uneven terrain when navigating through tilled fields. Such irregularities can lead to the robot drifting off course. To achieve effective autonomous navigation for agricultural robots, it is crucial to develop strategies for accurately detecting crop rows to mitigate the impact of external disturbances. In this paper, a novel algorithm of crop row detection for field robots that incorporates a multi-head attention mechanism and Res-Net model is presented (MRA-ResNet). First, real-time field video captured by an agricultural robot is processed for frame extraction to obtain a large number of images, which are initially convolved by an adaptive threshold filter to obtain key frame features. Then, the crop features in the field are learned by setting up a multi-region attention module, segmenting the crop and background, and dividing the learned crop feature images into several vertical image strips. These learned vertical image strips are then mapped to the low latitude for classification by vertical projection and MRA-ResNet. Finally, the center-line of the crop rows is further determined accurately. The experimental results show that MRA-ResNet can well realize crop row detection in the field.