Pseudo-label enhanced multi-feature fusion for millimeter-wave radar UAV detection

A weakly supervised learning method was investigated for unmanned aerial vehicle (UAV) detection based on millimeter-wave radar. To enhance feature representation, multiple categories of information were fused, including spectral energy distribution, motion trajectory variation, and statistical descriptors. Target pre-detection was conducted using a constant false alarm rate (CFAR) algorithm, which adaptively filtered the radar echo data and provided coarse localization. Doppler-related indexes were employed to reflect target motion parameters, and time-frequency domain information was used to derive energy characteristics for feature extraction. To further characterize the echo distribution, statistical measures like skewness, kurtosis, and energy ratio across close and far range bins were calculated. A LogitBoost ensemble classifier was employed to train the detection model by combining multiple weighted weak learners through iterative optimization. To address the scarcity of labeled samples, a pseudo-labeling strategy was introduced. The self-training mechanism automatically generated pseudo-labels from high-confidence predictions on unlabeled data and incorporated them into subsequent training cycles. Experimental validation was performed using a dataset comprising multiple real-world radar recordings with UAVs and empty field scenarios. The proposed method demonstrated robust performance under weak supervised conditions in three different scenarios. Compared with the baseline model, the final model achieved an improvement of approximately 7.2%/1.6%/6.6% in the area under the receiver operating characteristic curve (AUC) and a reduction of approximately 36.9%/0%/10.6% in the false alarm rate. Moreover, the model exhibited consistent detection accuracy in challenging environments with noise and background clutter. This study demonstrates that millimeter-wave radar detection performance for low-altitude, small-size UAVs can be significantly enhanced by combining multi-domain features and implementing pseudo-label augmentation techniques. The method provides practical value for real-time surveillance and airspace security applications.