A laser stripe extraction and segmentation method of windshield automatic assembly body in complex industrial scenes

In the automatic assembly process of automobile windshield, the use of laser stripe three-dimensional measurement technology to accurately and quickly complete the space posture measurement of the body is an important prerequisite for vision-guided assembly. Due to the complexity of the assembly environment, the laser stripes projected onto the fixed position of the vehicle inevitably suffer from noise pollution. In order to improve the influence of noise pollution on the accuracy of laser stripe segmentation, an improved U-net network training model based on attention mechanism and dilated convolution is proposed for laser stripe segmentation of vehicle body. In the structure of the traditional U-net, a CBAM attention module is introduced before each downsampling layer for a given feature map. Then dilated convolutions is used behind the pooling operation, and the dataset of body laser stripe images from the windshield automated assembly site is used for verification. The results show that the improved network model improves the accuracy, recall rate and F1 score and other evaluation indicators, and the average intersection and consolidation index is improved by 1.84% compared with the original U-net. Compared with the traditional vision algorithms, the improved U-net network has better detail retention ability in the stripe segmentation results, and improves the accuracy of laser stripe segmentation.