Mengxin Shi

Tracking multiple aerial targets within a group is a challenging task in the radar field. Group targets are usually closely spaced and have similar features, which leads to uncertain associations in target tracking. Additionally, targets within a group interact with each other, making it difficult to establish an accurate prior motion model. These factors limit the conventional model-driven tracking methods, resulting in false associations and tracking errors. Therefore, this paper proposes a group target tracking method based on spatial-temporal attention graph networks. First, a generalized graph representation is constructed to model the various interactions among group targets. Then, a two-stage network framework is introduced to learn the spatial and temporal relationships, thereby achieving robust tracking. Specifically, a message passing network with spatial-temporal attention is designed for data association of closely spaced targets, and a graph-aided filter state modification network is designed to correct the tracking error caused by mismatched motion models. Within the framework of multiple hypothesis tracking, data association and track filtering are optimized in a data-driven way. Simulation and experiment results demonstrate that the method outperforms the competing methods.

Ensuring constant false alarm rate (CFAR) detection in complex non-Gaussian heterogeneous backgrounds poses a major challenge in low-altitude target monitoring. Heterogeneous Gamma clutter is a typical non-Gaussian heterogeneous background. Its heavy-tailed characteristics, together with masking effects from strong neighboring scatter interference, cause inaccurate background level estimation in conventional detectors. As a result, the detection performance degrades significantly. To overcome above detection difficulties, this paper proposes a Gamma-distributed background Bayesian interference control CFAR (GBIC-CFAR) detector. A standard CFAR detection procedure is developed using Bayesian methods, resulting in a Bayesian predictive inference model for Gamma clutter. Meanwhile, by constructing a reasonable predictive inference distribution to control interference, effective compensation for interference is achieved. Additionally, to address errors caused by inaccurate prior knowledge of interference in the Bayesian detector, a method for selecting interference prior probabilities based on discrete truncated Gaussian function is proposed. By adjusting the truncation threshold, the prior probabilities are optimized, improving the robustness of the proposed detector. Finally, the effectiveness of the proposed algorithm is validated through simulations and real data.

Accurate estimation of targets' number is a crucial prerequisite for observing group target. Traditional methods estimate target number by resolving targets in the range/angle domain, but they become ineffective when targets are densely distributed or the system resolution is limited. While some super-resolution methods can estimate the number of densely distributed targets, they require the echo signals to remain stationary and suffer performance degradation when the group is highly dynamic and individuals are fast fluctuating. This paper proposes a target number estimation method using multi-angle incoherent observations. Firstly, this paper analyzes the fast fluctuation characteristics of echoes from a highly dynamic and densely distributed group target, and demonstrates the incoherence of echoes at different observation frames. Then the group target's multi-angle observation matrix and the corresponding covariance matrix is constructed. Finally, the relationship between the target number and the eigenvalue distribution of the observation covariance matrix is formulated, and the number is accurately estimated through the principal eigenvalues segmentation. Furthermore, by modifying the cost function of the principal eigenvalues segmentation method based on the Minimum Description Length criterion, this paper improves the “underestimation” issue of principal eigenvalues segmentation under low signal-to-noise ratio conditions. The proposed method can achieve the number estimation of a highly dynamic and densely distributed group target using only a single sensor with limited resolution. Simulation and experimental results demonstrate that the proposed method achieves a correct estimation rate of over $80 \%$ when the target spacing is less than the range resolution.

The robust tracking of group targets, such as bird flocks and UAV swarms, is crucial for situational awareness and individual tracking. The main characteristics of group targets are their complex spatial structures and highly dynamic behavior. Traditional single ellipsoidal tracking methods cannot accurately describe such complex shapes. On the other hand, existing multi-ellipsoidal tracking methods assume mutual independence between sub-objects. This assumption can lead to inaccurate responses to changes in group states and false associations between measurements and sub-objects. This paper proposes an adaptive multivariate approach for dynamic group target tracking. A multivariate model is first proposed to model sub-object states precisely. Both collective and internal structural relations are incorporated to adapt to the group's complicated dynamics. A variational Bayesian approach is derived to jointly estimate the multivariate state and the association weights of measurements. Additionally, an adaptive sub-object partitioning algorithm is proposed to analyze various group structures automatically. The effectiveness of the proposed method is verified by simulation and experiment results. The results demonstrate that our method outperforms state-of-the-art methods in tracking accuracy and robustness, especially in dynamic scenarios.