Robust Multi-Object Tracking Using Re-Identification Features and Graph Convolutional Networks

We propose a graph neural network-based framework for multi-object tracking that combines detection and association along with the use of a novel re-identification feature. We explore the combination of multiple appearance features within our framework to obtain a better representation and improve tracking accuracy. Data augmentations with random erase and random noise are utilized to improve robustness during tracking. We consider various types of losses during training, including a unique application of the triplet loss to improve overall network performance. Results are presented on the UAVDT benchmark dataset for aerial-based vehicle tracking under various conditions.