【摘要】 针对传统装配工序工时测定中，操作人员对手部抓取零件、对准装配位等动作细节时间点把控不足造成工序工时测定误差较大的问题，提出通过机器视觉中语义分割算法对人像进行提取的方法，以辅助操作人员更好地判断各工序的起止时间。首先，以工业装配场景视频为研究对象，选取人类装配视频数据集（HA-ViD）开展实验，采用MobileNetV3作为编码器提取多尺度特征；然后，在编码器与解码器的4个跳跃连接处嵌入双分支遮挡感知模块（DOA）-卷积注意力模块（CBAM），通过解码器经上采样与跳跃连接特征融合，完成装配动作精细化语义分割；最后，结合帧序列分析实现有效工时的自动统计。结果显示，该方法的单帧分割时间为32 ms，平均交并比（mIoU）达92.6%,Dice系数达93.3%，动作细节识别误差为0.14 s，且基于机器视觉辅助的工序工时测定更具有准确性。更多还原

【Abstract】 In traditional work hour measurement for assembly processes, significant errors often arise due to operators′ inadequate control over the precise timing of hand movements, such as grasping components and aligning them with assembly positions. To address this issue, a method is proposed that utilizes semantic segmentation algorithms in machine vision to extract human figures, thereby assisting operators in better determining the start and end times of each process. Firstly, industrial assembly scene videos are taken as the research objects, and experiments are conducted using the Human Assembly Video Dataset(HA-ViD). MobileNetV3 is employed as the encoder to extract multi-scale features. Then, a Dual-branch Occlusion-Aware(DOA)-Convolutional Block Attention Module(CBAM) is embedded at four skip connections between the encoder and decoder. Through upsampling by the decoder and feature fusion via skip connections, refined semantic segmentation of assembly actions is achieved. Finally, effective work hours are automatically counted by combining frame sequence analysis. The results show that this method achieves a single-frame segmentation time of 32 ms, with a mean Intersection over Union(mIoU) of 92.6%, a Dice coefficient of 93.3%, and an action detail recognition error of 0.14 s. Moreover, work hour measurement assisted by machine vision demonstrates greater accuracy.更多还原