基于Openpose改进的苹果生长方向检测

Apple growth direction detection based on improved Openpose

【目的】苹果采摘机器人在果园作业过程中,不仅需要精准地确定果实的位置,还需要对苹果的生长方向进行精准检测,这样才能够实现机械爪规划出仿人式采摘的抓取方式,实现果实的高效采摘。【方法】文章提出了一种基于改进Openpose的方法,实现了对自然生长状态下果实生长方向的精准检测。改进方案主要从两个方面进行,首选是通过ShuffleNet V2和CA注意力机制相融合的方式替换原VGG19结构,降低主干的参数量,然后是结合单个苹果关键点的特征对该模型进行了去除部分PAF分支的优化改进策略。【结果】改进后的Openpose在各个指标上,全面超过了原Openpose,其运行速度是改进前的6.56倍,对于mAP、m AP-s、mAP-b、AP50和AP75这些参数,分别增长9.18%、8.50%、11.56%、1.67%和6.35%,并且在各项指标上超越了AlphaPose和CFA算法。【结论】经过对Openpose模型改进前后的对比和现有算法对比,证明了该文算法在自然场景下对果实生长方向检测的优越性和鲁棒性。

[Purpose]The apple picking robot needs to accurately determine not only the position of the fruit but also the growth direction of the apple during the operation in the orchard,so that the mechanical claws can plan a human-like picking grasping method and realize the efficient picking of the fruit.[Method]The article proposes a method based on an improved Openpose to achieve accurate detection of the fruit growth direction in the natural growth state.The improvement scheme is mainly carried out in two aspects,the first choice is to replace the original VGG19 structure by the fusion of ShuffleNet V2 and CA attention mechanism to reduce the number of parameters of the backbone,and then the optimization improvement strategy of removing some PAF branches is carried out for this model by combining the characteristics of the key points of individual apples.[Result]The improved Openpose outperforms the original Openpose in all metrics,and its operation speed is 6.56 times faster than before the improvement,and for the parameters of mAP,mAP-s,mAP-b,AP50 and AP75,it increases by 9.18%,8.50%,11.56%,1.67%and 6.35%,respectively,and in all metrics outperformed the AlphaPose and CFA algorithms.[Conclusion]After comparing the Openpose before and after improvement with the existing algorithms,the superiority and robustness of this paper's algorithm for fruit growth direction detection in natural scenes are proved.