Answers are grounded in this paper and cite their sources.
Real-time pedestrian detection and multi-object tracking (MOT) in complex airport surveillance scenarios present significant challenges, including drastic scale variation, frequent occlusion, cluttered backgrounds, and diverse motion dynamics. Existing methods often fail to simultaneously satisfy accuracy, robustness, and real-time requirements in such demanding environments. To address these challenges, we propose AirMOTR, a unified detection-and-tracking framework tailored for airport-scene pedestrian perception. First, we integrate a lightweight two-dimensional Selective State Space Model (SS2D) into the backbone feature extractor to efficiently capture long-range spatial dependencies while maintaining low computational overhead. Second, we introduce three complementary feature enhancement modules: aFrequency-aware Interaction Block(FIB) that employs discrete wavelet transform to enrich feature representations in the frequency domain; aQuery-driven Context Extractor(QCE) that aggregates contextual semantics through detection-query guidance; and anAdaptive Spatial-Channel Gate(ASCG) that dynamically modulates feature responses across both spatial and channel dimensions. Third, we embed theMotion Mamba(MM) module within MM-Tracker, combining local correlation and global bidirectional scanning to generate accurate inter-frame motion maps. A novelMotion Margin Loss(MMLoss) is also proposed to adaptively adjust association boundaries based on per-target motion magnitude, improving tracking robustness for both slow-moving and fast-moving airport passengers. For training, we first pre-train on the large-scale CrowdHuman and WiderPedestrian datasets to strengthen pedestrian feature generalization, then fine-tune on our newly constructedAirportPassengerdataset and validate generalizability on the ASS (Airport Security Surveillance) dataset. Notably, SS2D andMMform a complementary pair: SS2D captures global spatial dependencies within individual frames, while MM leverages bidirectional scanning across temporal frames for motion modeling, jointly addressing the spatial and temporal challenges of airport surveillance. Extensive experiments demonstrate that AirMOTR outperforms state-of-the-art methods on all benchmarks while maintaining real-time inference speed (≈41.6 FPS), providing a robust and practical solution for intelligent airport surveillance. Code is available at https://github.com/zhoucancqa/AirMOTR.
No takes yet. Share an insight, caveat, or question.
Can et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: