Contrastive Multi-Modal Fusion for Enhanced Airport Surface Surveillance

Remote and Virtual Tower (RVT) is increasingly attracting attention for its potential to reduce human resource requirements and construction costs in air traffic management. While advances in artificial intelligence have enhanced the safety and efficiency of RVT systems, key tasks such as aircraft detection remain challenged by limitations in the data quality of single modality, affecting the accuracy of current implementations. To address these challenges, we introduce a framework named Vision-ADSB Network (VAD-Net) to fuse primary surface monitoring data from RVT. Unlike traditional methods, VAD-Net integrates visual data with Automatic Dependent Surveillance Broadcast (ADS-B) information using a specialized method based on contrastive learning, generating a comprehensive semantic representation. By fusing camera and ADS-B data, VAD-Net addresses the shortcomings of single-modality data and enhances the overall monitoring capabilities. Experimental results, validated with data collected from RVT systems, demonstrate that the VAD-Net model outperforms existing methods, even when working with limited datasets.