A multiple CNN-GRU (MCNN-GRU) collision warning network model is proposed to improve the transfer safety of ship surface targets. The network combines the feature extraction ability of convolutional neural network (CNN) for single time step information and the memory ability of gated recurrent unit (GRU) for time sequence. Multi-channel network structure is used to improve the processing performance of multi-time step information features. Target detection network, key point detection network, pose calculation model and collision detection method are used to produce ship surface target collision warning data set. The experimental results on different network data sets show that the collision warning accuracy of the model is 92. 44%, and it has a good effect.