Abstract: To tackle the issue of previous track fusion algorithms′ heavy reliance on prior information, a track fusion algorithm based on fully convolutional network with multi-order difference loss is proposed. The various local tracks are subjected to spatiotemporal registration and track association at the fusion center. Through the design of the fully convolutional structure, the problems of large amount of parameters and difficult training caused by the use of fully connected layers in the traditional convolutional neural network model are avoided. Higher-precision fusion results are obtained by calculating the weighted loss of the track and its first- and second-order differences. The ablation experiments show that the track fusion algorithm proposed in this paper has a small model, strong convergence, high precision and moderate computing time. Simulation results confirmed that the algorithm proposed in this paper does not require any previous information. The proposed algorithm's fusion accuracy is better than the variance weighted fusion algorithm and the Kalman filter fusion algorithm when the noise parameters cannot be accurately estimated. The experimental results confirm the effectiveness and feasibility of the algorithm proposed in this paper.