Field & Road-CENet: Semantic Segmentation for Agricultural Machinery Trajectory Segmentation

Abstract

Agricultural machinery management is the key to agricultural production, andtrajectory segmentation lays an important foundation for machinerymanagement. For big data platform of agricultural machinery, it is hoped tosimultaneously improve both of segmentation accuracy and segmentationefficiency to satisfy the processing requirements. However, traditional machinelearning algorithms need to manually adjust parameters and do not integratemultiple features when dealing with trajectory segmentation. This paper aims atthe above problems by modifying the CE-Net model to improve the segmentationaccuracy and ensure the processing efficiency. A creative method for constructingtrajectory image from discrete trajectory data was developed in this paper. Thenew trajectory image showed both the temporal and the spatial characteristics ofthe operation. Then, a semantic segmentation algorithm, Field & Road- CENet,was put forward. The proposed algorithm modified the network CE-Net by addingtwo modules, Standard Convolution Residual Block and Global MaxpoolingAttention Mechanism, to fuse original feature information and enhance thesemantic expression of low-level feature maps. Two cotton sowing datasets werebuilt, including the meter level and the centimeter level. Experiment results showthat Field & Road-CENet performed well on both datasets. In the fieldsegmentation that is the most concerned, the identification accuracy reached97.8% and 95.2%, respectively, and the average accuracy of field and road was94.2% and 88.3%, respectively. In conclusion, this work verifies the feasibility ofusing semantic segmentation to realize trajectory segmentation of agriculturalmachinery. Compared with the current researches, the proposed method isapplicable to trajectory data with two precisions, which has stronger domaingeneralization ability. And it performs quite fast with an average inference time of 0.044 s for each image block, demonstrating that the proposed algorithm issuitable for the big data processing of agricultural machinery.