针对无人机机载毫米波雷达对“低慢小”目标跟踪中存在的量测噪声时变、非高斯离群点干扰，以及高机动场景下模型切换滞后等问题，提出一种基于毫米波雷达点云统计特性的自适应交互多模型无迹卡尔曼滤波(Adaptive Interacting Multiple Model - Unscented Kalman Filter，ADIMM-UKF)无人机目标跟踪算法。该方法以匀速(CV)、匀加速(CA)与转弯(CT)三模型IMM-UKF为框架，利用点云空间分布特性构建自适应量测噪声协方差矩阵，结合绝对中位差(Median Absolute Deviation，MAD)准则剔除异常散射点，提升观测数据质量与表征精度；同时在UKF子滤波器中引入Student-t重尾分布建模，对新息进行鲁棒加权更新，抑制离群量测对状态估计的干扰；设计基于Sigmoid非线性映射的瞬态激活机制，在线修正模型转移概率矩阵，提升突发机动下的模型切换速率与匹配精度。实验分别基于雷达目标模拟器数据和真实无人机点云数据开展验证，结果表明：所提算法在二维模拟场景下的位置、速度均方根误差相较于AIMM-UKF分别降低18.7%和14.3%，在三维实测场景下的位置、速度误差分别降低22.9%和12.4%，验证了算法在实际应用环境中的有效性与鲁棒性。

To address the problems of time-varying measurement noise, non-Gaussian outlier interference, and model-switching delay in tracking low-altitude, slow-speed, and small targets with airborne millimeter-wave radar, this paper proposes an Adaptive Interacting Multiple Model–Unscented Kalman Filter (ADIMM-UKF) algorithm based on the statistical characteristics of radar point clouds. Under a three-model IMM-UKF framework including constant velocity (CV), constant acceleration (CA), and coordinated turn (CT) models, an adaptive measurement noise covariance matrix is constructed from the spatial distribution of point clouds, while the Median Absolute Deviation (MAD) criterion is used to remove abnormal scattering points, thereby improving measurement quality. In addition, a Student’s t-distribution is introduced into the UKF sub-filters for robust innovation weighting, reducing the influence of outlier measurements on state estimation. A transient activation mechanism based on Sigmoid nonlinear mapping is further designed to update the model transition probability matrix online, improving model-switching responsiveness and matching accuracy under abrupt maneuvers. Experiments using both radar target simulator data and real UAV point-cloud data show that, compared with AIMM-UKF, the proposed algorithm reduces the root mean square errors of position and velocity by 18.7% and 14.3% in a two-dimensional simulated scenario, and by 22.9% and 12.4% in a three-dimensional real-world scenario, respectively. These results demonstrate the effectiveness and robustness of the proposed algorithm in practical applications.