Multi-sensor track association and interrupted track association based on deep learning algorithms
Songtao Hu, Liang Chen, Jun Yang · IET conference proceedings. · 2024
As the radar detection environment of modern battlefield becomes more and more complex, the limitations of traditional radar multi-sensor correlation methods are gradually revealed. The traditional method of radar track correlation is difficult to deal with the environment with high target density and maintain high accuracy. The traditional method of interrupting track correlation is weak in maneuverability and depends on the construction of motion model. To solve these problems, a track association model based on twin neural network algorithm and an interrupted track association model based on depth time contrast algorithm are proposed in this study. The track association model uses bidirectional recurrent neural network with shared weights and interactive attention mechanism to process the feature information to obtain similarity vector, so as to realize the judgment of track association. The interrupted track association model, by extracting the time sequence feature vector of the old and new track, and calculating the similarity of the track depth feature vector, can obtain the similar situation vector of the old and new track, and then realize the correlation judgment of the interrupted track. The accuracy of track correlation model is 98.7%; The average error correlation rate of interrupted track association model is 0.57%. The average missing correlation ra te was 0.1%. The results of this study have certain value to the field of track association, and can be used as a technical reference for multi-sensor track association.