关键词: 多目标关联与跟踪, 混合特征, 关联不确定, 高斯混合模型, 协同态势感知

Abstract: Cooperative perception based on multi-UAV passive detection can enhance covert detection and continuous tracking of multiple targets in complex environments. However, incomplete 2D observations from passive sensors and weak target appearance features lead to key challenges in passive multi-target association and tracking, including association feature scarcity, a high risk of mis-association, and difficulty in 3D state estimation. To address these bottlenecks, this paper proposes a hybrid feature-based passive multi-target data association and tracking method. First, to tackle association feature scarcity under passive observation and the degradation of a single association criterion, a hybrid feature criterion is constructed by integrating 2D point topological distribution, trajectory motion characteristics, and spatiotemporal geometric constraints. The criterion weights are adaptively adjusted according to the sensor geometric configuration, thereby achieving complementary enhancement of multiple weak association cues and improving the accuracy and robustness of inter-sensor target association. Second, to address the difficulty of 3D multi-target state estimation caused by mis-association, the Gaussian Mixture Model is employed to characterize the non-Gaussian distribution of fused measurement errors from coarse localization under association uncertainty. The GMM-based measurement model is formulated as an observation factor and incorporated into a factor graph optimization-based multi-target tracking framework. Within a sliding window, observation factors, state transition factors, and target mutual exclusion factors are jointly optimized to achieve robust estimation and stable tracking of multi-target states. Simulation results show that the proposed method achieves an association precision of 90.7% in complex scenarios. Under multi-target random motion mode, positioning RMSE is reduced by 34.3% and 47.7% compared to fusion-based tracking and filtering-based tracking methods. The multi-target tracking accuracy MOTA is improved by 44.3% and 57.2%. Real-world indoor experiments further verify the engineering applicability and tracking stability of the method. The results provide an effective technical path for multi-target passive detection cooperative perception of distributed aerial vehicles.

Key words: multi-target association and tracking, hybrid features, association uncertainty, Gaussian Mixture Model, cooperative perception