桥式起重机部件多任务学习Mask R-CNN分割与关键点识别方法

Multi-task Learning Mask R-CNN Segmentation and Key Point Recognition Method for Bridge Crane Components

起重机运行状态实时检测是工业安全生产的重要保障之一。针对起重机部件提出基于多任务学习Mask R-CNN的分割与关键点网络结构,该结构由Mask R-CNN定位框与区域语义分割网络、DeepLabCut关键点检测网络构成;在吊钩桥式起重机中采集数据进行测试,利用贪婪线性搜索算法与贝叶斯优化算法,搜索得到此模型的最优超参数组合为:学习率0.005,批数2,学习率策略为余弦衰减。该模型测试误差为2.46个像素点,测试AP可达95%,像素点误差反映到实际误差在5 cm以内,满足实际检测需求,可拓展应用于无人化、自动化起重机运行状态监测。

Real time detection of crane operation status is one of the important guarantees for industrial safety production.A segmentation and key point network structure based on multi task learning mask R-CNN is proposed R-CNN crane key parts positioning frame and regional semantic segmentation network,crane key point detection network based on DeepLabCut;in a hook crane to collect data for testing,using greedy linear search algorithm and Bayesian optimization algorithm,the optimal super parameter combination of this model is:learning rate 0.005,batch size 2,learning rate strategy is cosine decay.The test error of the model is 2.46 pixels,and the test AP is up to 95%.The pixel error can be reflected within 5 cm of the actual error,which can meet the actual detection needs,and can be extended to the unmanned and automatic crane operation state monitoring.