This paper presents a track initiation method of radar target based on deep learning. The algorithm transforms track initiation into the problem of neural networks for classification. The two classifications are real track and false track respectively. Firstly, associate target plots after spatial registration by Annular gate roughly, and get the temporary track. Secondly, the input vector of neural network is generated by eigenvector modeling of temporary track. Then extract neural network training samples from simulation system. And design structure of deep full connection neural network. Then get optimized model parameters by training the neural network model. Finally, using the optimized model parameters calculate initial track online. The simulation result validate the efficiency of our algorithm.