Precise UAV MMW-Vision Positioning: A Modal-Oriented Self-Tuning Fusion Framework

Precise real-time unmanned aerial vehicle (UAV) positioning is crucial for preventing unauthorized UAVs from damaging cooperative intelligent transportation systems (C-ITSs). However, UAV positioning remains extremely challenging due to the small target size and high flexibility. Therefore, we develop a modal-oriented self-tuning fusion framework for precise UAV millimeter-wave(MMW)-vision positioning. The framework selects and extracts cross-modal features based on modality characters, and migrates the Doppler features of MMW radar data to the image features for precise pixel-level positioning. Based on the framework, a modal-oriented self-tuning fusion network is proposed to adaptively enhance UAV feature without direct supervision by exploiting the cross-modal correlations. A novel characteristic-based 3DMMW feature extraction method is presented to extract UAV Doppler motion characteristics while a self-tuning cross-modal affine transfer is proposed for UAV visual feature enhancement. Due to lack of dataset for our task, we establish a practical positioning platform and two novel datasets containing synchronized visual images and MMW radio frequency (RF) sequences in various scenarios. Experimental results confirm that our framework outperforms the benchmark methods in terms of positioning accuracy while maintaining real-time performance. Moreover, ablation studies also confirm the effectiveness of each module in the framework.