UAV Trajectory Tracking via RNN-enhanced IMM-KF with ADS-B Data

UAV Trajectory Tracking via RNN-enhanced IMM-KF with ADS-B Data

Yian Zhu, Ziye Jia, Qihui Wu, Chao Dong, Zirui Zhuang, Huiling Hu, Qi Cai

With the increasing use of autonomous unmanned aerial vehicles (UAVs), it is critical to ensure that they are continuously tracked and controlled, especially when UAVs operate beyond the communication range of ground stations (GSs). Conventional surveillance methods for UAVs, such as satellite communications, ground mobile networks and radars are subject to high costs and latency.

The automatic dependent surveillance-broadcast (ADS-B) emerges as a promising method to monitor UAVs, due to the advantages of real-time capabilities, easy deployment and affordable cost. Therefore, we employ the ADS-B for UAV trajectory tracking in this work. However, the inherent noise in the transmitted data poses an obstacle for precisely tracking UAVs. Hence, we propose the algorithm of recurrent neural network-enhanced interacting multiple model-Kalman filter (RNN-enhanced IMM-KF) for UAV trajectory filtering.

Specifically, the algorithm utilizes the RNN to capture the maneuvering behavior of UAVs and the noise level in the ADS-B data. Moreover, accurate UAV tracking is achieved by adaptively adjusting the process noise matrix and observation noise matrix of IMM-KF with the assistance of the RNN. The proposed algorithm can facilitate GSs to make timely decisions during trajectory deviations of UAVs and improve the airspace safety.

Finally, via comprehensive simulations, the total root mean square error of the proposed algorithm decreases by 28.56%, compared to the traditional IMM-KF.

Subjects:	Systems and Control (eess.SY)
Cite as:	arXiv:2312.15721 [eess.SY]
	(or arXiv:2312.15721v1 [eess.SY] for this version)

Submission history

From: Yian Zhu [view email]
[v1] Mon, 25 Dec 2023 13:14:13 UTC (11,567 KB)

 Summary

• The paper proposes an RNN-enhanced IMM-KF algorithm for tracking the trajectories of UAVs using ADS-B data. ADS-B allows UAVs to broadcast their position and velocity in real-time, but the data contains noise.
• Two dynamic models are defined - a constant velocity (CV) model for stable cruising flight, and a constant jerk (CJ) model for high-agility maneuvering flight. An IMM filter is used to switch between these models.
• An RNN is designed to dynamically adjust the process noise covariance matrix Q and observation noise covariance matrix R in the IMM-KF filter. This helps deal with changing noise levels and model inaccuracies.
• The input to the RNN at each timestep is the difference between the actual observation and the IMM-KF position estimate. The RNN outputs estimated parameters for Q and R.
• In simulations, the proposed RNN-enhanced IMM-KF reduces the root mean square error by 28.56% compared to standard IMM-KF with fixed noise matrices. It prevents filter divergence.
• The approach can enable ground stations to accurately track UAV trajectories in real-time using ADS-B, even when noise is present or UAVs are maneuvering. This improves airspace safety and management.

Authors

Based on the author information in the paper:

• The authors are from institutions in China:

1. Nanjing University of Aeronautics and Astronautics
2. Beijing University of Posts and Telecommunications
3. Middle-south Regional Air Traffic Management Bureau of CAAC

• The corresponding author is Qihui Wu from Nanjing University of Aeronautics and Astronautics. His email is listed as wuqihui@nuaa.edu.cn.
• Two of the authors seem to have previous related work in this area:

1. Ziye Jia - has published papers on UAV trajectory prediction using ADS-B data and wireless networking for UAVs

2. Qihui Wu - has previous published works on tracking of maneuvering targets and air traffic management

So the authors seem to have experience in UAV communications, tracking, and air traffic management. Their institutional affiliations are with universities and air traffic control organizations in China.

It does not appear that the authors have released any code or data artifacts for this particular RNN-enhanced IMM-KF method yet. Sharing such artifacts could be useful for validation and extension of the work by others. But the lack of artifacts currently limits reproduciblity.

Resources Used

The paper does not provide specific details about the software, languages, or hardware used for the simulations and code implementation. However, some clues can be inferred:

• They performed simulations using a fixed-wing UAV model from the MATLAB toolbox. So MATLAB was likely used as the main programming environment.
• The proposed algorithm involves Kalman filtering and recurrent neural networks (RNNs). These techniques are commonly implemented using Python libraries like NumPy, SciPy, TensorFlow, Keras, PyTorch, etc. So Python was likely used, in addition to or instead of MATLAB.
• The RNN model uses Long Short-Term Memory (LSTM) cells. LSTMs are a standard RNN architecture well-suited for sequence data like trajectories.
• The simulations involved generating 3D position and velocity data. This type of vectorized data is easy to handle in MATLAB and Python using multidimensional arrays.

So in summary, the simulations were likely developed in MATLAB and/or Python, using common libraries like TensorFlow, on typical desktop/laptop hardware. But the paper does not provide direct confirmation of the specific software, languages, or hardware resources used.