Khái niệm cốt lõi
The paper proposes a federated radar network model that enables distributed point cloud processing and scene reconstruction without exchanging raw sensor data. The federated approach outperforms classical cooperation in terms of communication overhead while maintaining comparable accuracy for multi-target tracking.
Tóm tắt
The paper investigates two distributed radar processing models - cooperation and federation - for enhancing the sensing and tracking capabilities of a network of resource-constrained MIMO radars.
In the cooperation model, the radars exchange pre-processed point cloud (PC) data to jointly reconstruct a global posterior probability measure of the observed scene. This allows them to mitigate shadowing effects and improve resolution compared to individual radars.
The paper then introduces a federated radar model, where the radars share the parameters of a local Bayesian posterior measure instead of raw PC data. This federated approach reduces the communication overhead by 20-25 times compared to cooperation, while maintaining comparable accuracy for multi-target tracking.
The key aspects are:
Radars operate independently to extract local 3D PC information, but cooperate by sharing information over a sidelink communication channel.
In the cooperation model, radars exchange pre-processed PC data to jointly reconstruct a global posterior probability measure.
In the federated model, radars share the parameters of a local Bayesian posterior measure, rather than raw PC data, to reconstruct a federated posterior.
Experiments with a real-time demonstration platform show that the federated approach is less sensitive to unresolved targets and has significantly lower communication requirements, while the cooperation model provides slightly better average localization accuracy.
The paper also discusses the use of Gaussian mixture models to represent the local and federated posterior probability measures, and the role of priors based on target motion models.
Thống kê
The mean absolute error (MAE) for target localization is around 0.09 m for the cooperation model and 0.11 m for the federated model.
The probability of unresolved targets is around 1% for the cooperation and federated models, compared to 10% for the isolated radar.
The communication overhead for the federated model is 20-25 times lower than the cooperation model.
Trích dẫn
"Federation makes minimal use of the sidelink communication channel (20 ÷ 25 times lower bandwidth use) and is less sensitive to unresolved targets. On the other hand, cooperation reduces the mean absolute target estimation error of about 20%."