摘要
Abstract
To address the issue of low monocular vision pose tracking accuracy for mobile robots in greenhouse environ-ments,an improved Monodepth2-based model is proposed.First,a monocular image dataset of greenhouse strawberries with depth and pose ground truth is constructed.Then,based on Monodepth2 as the baseline,the backbone network is replaced with ResNext50,a hybrid convolution module is designed,and a coordinate attention mechanism is introduced to form the Mono-HCCA model.Finally,Mono-HCCA model serves as the teacher model,and a lightweight student model,Mono-LW,is obtained through knowledge distillation.Experimental results show that the root mean square er-rors for position and pose estimation in Mono-HCCA are 0.144 m and 0.132 rad,respectively,representing reductions of 29.1％and 20.5％compared to the original model.The student model Mono-LW achieves a frame rate of 50 fps,2.3 times that of the teacher model,while its position error increases by only 4.9％compared to the teacher model.In summary,the Mono-LW model maintains high pose tracking accuracy while meeting real-time requirements,provi-ding effective technical support for autonomous navigation of greenhouse mobile robots in complex environments.关键词
温室移动机器人/单目视觉/位姿跟踪/Monodepth2/知识蒸馏Key words
greenhouse mobile robots/monocular vision/pose tracking/Monodepth2/knowledge distillation分类
农业科技
