Radar Target Track Before Detect Based on Sequence-to-Sequence Deep Neural Network

In response to the problems of high computational complexity and poor adaptability to complex scenes in traditional TBD methods for detecting stealth small targets, this paper proposes radar target track before detect based on sequence-to-sequence deep neural network. Firstly, this paper studies multidimensional feature extraction methods, including time amplitude, time phase, time Doppler, etc. Then, a sequence-to sequence network model is designed and corresponding principle is detailed described. Finally, considering the performance of deep models heavily relies on a large number of samples, this paper proposes a solution of sample annotation for practical radar detection scenarios. The real-data test shows that the method proposed in this paper can significantly improve the tracking distance of the small targets at long distances, and can effectively suppress false short tracks, laying a good foundation for engineering application of the algorithm.