Graph-Based Target Association for Multi-Drone Collaborative Perception Under Imperfect Detection Conditions

Multi-drone collaborative perception aims to address single-drone viewpoint limitations. The existing matching and association methods based on visual features and spatial topology rely heavily on detection, making it challenging to associate targets under imperfect detection conditions. To address this issue, a Graph-Based Target Association Network (GTA-Net) is proposed to utilize graph matching to associate the key objects before affine transforming and matching both detected and undetected targets. The Key Object Detection Network (KODN) finds the key object that is more likely to be a True Positive and more important. The Graph Feature Network (GFN) treats the key objects as graph nodes and extracts the graph feature. The Association Module utilizes graph matching to associate the top-k-like matching objects and iterable affine transformation to associate all objects. The experiment results show that our method achieved a 42% accuracy improvement on the public dataset. The ablation experiments under imperfect detection simulation demonstrate robust performance.