Objective In urban infrared low-altitude security monitoring scenarios, micro-small intrusive unmanned aerial vehicles (UAVs) and flying birds produce highly coupled imaging signatures and induce prominent false-positive and false-negative classification errors. Constrained by infrared imaging mechanisms, both targets behave as weak-texture small objects under long-distance observation. Noise masks valid image details, blurs target edges and degrades contour features to increase discrimination difficulty. Bird-triggered false alarms frequently activate counter-measure devices, consume substantial response resources and may cause incorrect counter-decisions. Missed detections of intrusive UAVs disable security defence lines and bring tangible air-space safety risks. Available UAV-bird discrimination techniques fall into two technical categories: conventional trajectory-based analysis and temporal deep-learning methods. Traditional trajectory-driven approaches employ hand-crafted motion statistics such as velocity and climb rate. They fail to exploit long-term temporal correlations hidden inside motion sequences. Recognition performance drops significantly under temporary target occlusion and tracking drift that generate short and fragmented tracks. Deep spatio-temporal models deliver powerful automatic high-dimensional feature-mining capacity, yet large parameter sizes and heavy inference loads impose high hardware-computation demands and break real-time processing constraints for low-power infrared embedded terminals. Targeted at these practical engineering bottlenecks, this study investigates lightweight UAV-bird classification algorithms for infrared image sequences based on inherent physical-morphology discrepancies between rigid multi-rotor-UAV structures and flexible bird biological bodies. Balanced recognition accuracy and computational efficiency achieve high-precision target discrimination and satisfy embedded-system deployment requirements.

Methods A dual-feature decision-level fusion classification algorithm based on D-S evidence theory is proposed. Temporal-deformation information and spatial contour information are mined from target physical essences to realize target discrimination with dual physical features. For temporal area-fluctuation features, pixel-area time-series are constructed from successive frames. Least-squares linear fitting is applied to area sequences. Residual statistical analysis and fast Fourier transform operate on fitting residuals to quantify periodic area-oscillation signatures induced by bird wing-flapping behaviours. These signatures characterize autonomous deformation from biological wing opening-closing. For posture-adaptive contour morphological features, pinhole imaging model, target centroid coordinates and sliding-filtered pixel-area sequences are combined. Sliding-filter processing removes high-frequency area disturbance from wing-flapping and mitigates deformation-induced interference in depth estimation. Relative three-dimensional target trajectories are reconstructed. These trajectories only reflect relative motion trends without true physical-scale data and fit application conditions of monocular infrared cameras. Fixed contour-analysis strategies cannot adapt to variable observation perspectives. Three typical flight attitudes are identified from reconstructed trajectories: far-near flight, horizontal flight and oblique flight. Posture-adaptive contour analysis is executed for each flight attitude to extract multi-dimensional morphological descriptors including wing protrusion degree, contour overlap ratio and wing-flapping synchronization index. D-S evidence theory implements decision-level fusion for temporal area features and contour morphological features. Complementary information from two feature sets relieves discrimination errors caused by single-feature failure from noise or attitude mutation and generates final classification outputs.

Results and Discussions Validation experiments are conducted on the public infrared dataset Drone-detection-dataset. The proposed algorithm achieves an F₁-Score of 96.51%, outperforming the optimal baseline algorithm by 5.78%. Periodic area oscillations from bird wing-flapping serve as valid discriminative cues for UAV-bird differentiation. Fixed mechanical structures of multi-rotor UAVs rule out autonomous biological-body deformation and produce smooth area-variation curves with small-amplitude fluctuations. Target-camera distance changes dominate UAV pixel-area variations. Periodic expansion and contraction of bird wings produce large-magnitude pixel-area mutations independent of imaging distances. The posture-adaptive mechanism supports contour analysis under far-near flight, horizontal flight and oblique flight. Bird-wing visible ranges and swing behaviours differ across these attitudes, and noticeable intrinsic contour discrepancies exist between birds and UAVs. Differentiated morphological metrics are extracted for target discrimination. Unlike deep-learning-based alternatives, this physical-feature-driven recognition scheme requires no large-scale annotated datasets for training and learns no hidden features from sample labels. It reduces data-collection and annotation costs for practical engineering. Complex real-world field environments bring interference challenges. Flying insects are widespread in natural scenes and constitute typical positive-false-alarm sources. High-frequency insect wing-beating generates high-frequency periodic area oscillations and bird-similar feature responses. Flapping-wing bionic UAVs are absent from the adopted dataset. Such UAVs imitate bird-flapping kinematics and produce periodic area and contour variations. Nevertheless, the dual-feature framework can partially separate flapping-wing UAVs from birds. Extremely low equivalent thermal-radiation intensity over wing regions of certain flapping-wing UAVs prevents valid observable wing contours under infrared imaging, and only fuselage regions provide prominent identifiable characteristics.

Conclusions A lightweight dual-feature fusion algorithm realizes UAV-bird recognition for infrared image sequences. Temporal area-fluctuation features and posture-adaptive contour morphological features are fused at the decision level via D-S evidence theory using essential physical differences between rigid multi-rotor-UAV airframes and flexible bird bodies. The algorithm relies purely on physical-morphology and temporal-variation features and avoids complicated network operations with controllable overall computation loads. Competitive classification performance is obtained without heavy deep-learning-related computational overhead. It matches hardware specifications of low-power infrared embedded terminals and delivers feasible technical support for urban low-altitude airspace governance and early warning against illegal multi-rotor-UAV intrusions. Limited dataset coverage leaves generalization performance insufficiently verified for special targets such as bionic flapping-wing UAVs. Degraded contour extraction under long-distance weak targets, strong-noise and low-SNR conditions impairs morphological-metric outputs, and feature-extraction stability needs further improvement. Dedicated studies will focus on discrimination between flapping-wing bionic UAVs and birds. Algorithm frameworks will be extended for multi-modal surveillance-data compatibility.