Multi-model KalmanNet for maneuvering target tracking

作者：Xuehan Han, Ling Ding, Peng Cheng, WenWen Zeng, Xin Zhang, Wen Zheng, Zheng Le · 发表于：IET conference proceedings. · 年份：2024 · DOI：10.1049/icp.2024.1112 · 被引用次数：2 · 研究领域：Target Tracking and Data Fusion in Sensor Networks、Neural Networks and Applications

The performance of traditional target tracking algorithms usually depends on the selection of preset motion models. When the real motion model of the target is mismatched with the preset model, the performance of the filter will degrade. How to dynamically update motion models to mitigate performance degradation caused by model mismatches has become an important issue. In this paper, we propose Multi-model KalmanNet, which is a multi-model network based on KalmanNet and can dynamically update motion models. While retaining the Kalman filter framework, it can learn the similarity between the target's motion model and the network's preset motion models, and uses the similarity as the weight of each preset model. By continuously updating the weights, it can alleviate the filtering performance degradation caused by model mismatch.