What question did this study set out to answer?

To develop an adaptive algorithm that improves target tracking accuracy for high manoeuvring objects by effectively handling model switching and noise suppression.

January 23, 2026Open Access

Adaptive High Manoeuvring Target Tracking Algorithm Based on CNN‐LSTM Fusion Architecture

Key Points

To develop an adaptive algorithm that improves target tracking accuracy for high manoeuvring objects by effectively handling model switching and noise suppression.
Developed an adaptive interacting multiple model unscented Kalman filter (IMM-UKF) algorithm.
Utilized a convolutional neural network (CNN) to extract spatial features and a long short-term memory network (LSTM) to capture temporal dynamics.
Constructed a multi-dimensional feature space from state estimation, observation data, and model probabilities.
Introduced a sliding average buffer mechanism to smooth predictions.
Achieved a reduction in position RMSE by 11.2% in periodic motion and 14.87% in compound manoeuvring scenarios.
Achieved a reduction in velocity RMSE by 19.96% in periodic motion and 21.50% in compound manoeuvring scenarios.
Improved model switching accuracy and enhanced tracking performance under high-manoeuvrability conditions.

Abstract

ABSTRACT High manoeuvring target tracking remains a challenging problem due to complex motion patterns, frequent model switching and strong nonlinear relationships between system states and observations. The interacting multiple model (IMM) algorithm integrated with the unscented Kalman filter (UKF) is widely used for such scenarios. However, its performance is significantly limited by reliance on a fixed or heuristically designed transition probability matrix (TPM), which leads to model switching lag and degradation in tracking accuracy during abrupt manoeuvres. Moreover, existing deep learning‐assisted IMM methods often fail to effectively fuse spatiotemporal features and suppress noise. To solve these gaps, an adaptive interacting multiple model unscented Kalman filter (IMM‐UKF) algorithm based on a convolutional neural network and long short‐term memory network (CNN‐LSTM) fusion architecture is proposed. A multi‐dimensional feature space including state estimation, observation data and model probability is constructed by the algorithm, which is then used as input to the neural network model. Latent spatial features are extracted by the CNN module and subsequently processed by the LSTM network to capture temporal dynamic characteristics, which achieve real‐time dynamic estimation and adaptive optimisation of the model TPM. In addition, a sliding average buffer mechanism is introduced to smooth the prediction outputs and reduce the impact of disturbances on estimation performance. Simulation results show that the proposed algorithm outperforms the IMM‐UKF algorithm. In the periodic motion scenario, the proposed algorithm reduces position and velocity root mean square error (RMSE) by 11.2% and 19.96%, respectively. In the compound manoeuvring motion scenario, the proposed algorithm reduces the position and velocity RMSE by 14.87% and 21.50%, respectively. The proposed algorithm effectively improves model switching accuracy, increases the probability of matching the dominant model and significantly enhances tracking performance under high‐manoeuvrability conditions.

Read Full Paperexternally

Bookmark

View Full Paper

Cite This Study

Cui et al. (Thu,) studied this question.

synapsesocial.com/papers/69730ef2c8125b09b0d1eb99 https://doi.org/https://doi.org/10.1049/rsn2.70115

Bookmark

View Full Paper