همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Selective Cross-Feature Fusion Network (SCFF-Net) for Vehicle Re-Identification in Urban Surveillance

نویسندگان :

Morteza Makhloughi¹ Ali Amiri² Aydin Kaya³

1- دانشگاه زنجان 2- دانشگاه زنجان 3- Hacettepe University

کلمات کلیدی :

Vehicle Re-Identification،Hybrid CNN–Transformer Network،Selective Feature Fusion،Intelligent Transportation Systems (ITS)،Urban Surveillance

چکیده :

Vehicle re-identification (Re-ID) plays an increasingly important role in intelligent transportation systems (ITS) and urban surveillance by enabling vehicle tracking across multiple cameras with non-overlapping views. The task is challenging due to dramatic appearance variations caused by illumination, occlusion, background clutter, and camera viewpoints.Most deep-learning-based methods focus either on local part-level features or global semantic embeddings, leading to redundancy and loss of fine detail. To address these issues, this paper presents the Selective Cross-Feature Fusion Network (SCFF-Net) — a lightweight hybrid CNN–Transformer architecture that adaptively integrates convolutional texture cues and transformer-based global semantics. The key innovation is the Selective Cross-Feature Fusion (SCFF) module, which performs adaptive gating to filter redundant activations and emphasize complementary feature responses. Unlike previous hybrid methods with static fusion, SCFF-Net learns dynamic, input-dependent fusion weights that enhance discriminative representations. Extensive experiments on two public benchmarks, VeRi-776 and VehicleID, demonstrate that SCFF-Net achieves superior performance (85.2% mAP / 97.9% Rank-1 on VeRi-776 and 79.4% mAP on VehicleID) while maintaining real-time inference efficiency (≈ 8 ms/image). In addition, cross-domain experiments confirm its generalization capacity to unseen environments. With a balanced trade-off between accuracy and computational cost, SCFF-Net offers a practical and scalable solution for city-scale vehicle monitoring.