In order to realize collision prediction in the process of loading and unloading of robot

vision technology is used to segment the image captured by the workpiece to determine whether there will be interference. Firstly

the distance sensor is used to realize one-dimensional positioning

after that the image is collected. To predict the collisions to come

the image semantic segmentation method based on the combination of full convolution network and conditional random field is proposed. In this method

VGG-16 is used as the basic network

and the remote residual connection method of residual network is used to optimize the full convolution network

and the back-end connection conditions are used to refine the segmentation results. With deep learning framework advanced by Keras

after training the algorithm model

the average pixel accuracy of workpiece grabbing image segmentation is 99. 17 %

the average intersection ratio is 85. 32 %

and the average time it takes is 0. 86 s

which reached the expectation of the quality and real time requirement of image segmentation. The accuracy rate of robot collision detection is 99. 96 %

which can prove advance of the method in this paper.