Data Association with Graph Network for Multi-Object Tracking

Yubin Wu, Hao Sheng, Shuai Wang, Yang Liu, Wei Ke, Zhang Xiong

Abstract: Multi-Object Tracking (MOT) methods within Tracking-by-Detection paradigm are usually modeled as graph problem. It is challenging to associate objects in dense scenes with frequent occlusion. To further model object interactions and repair detection errors, we use graph network to extract embeddings for data association. Graph neural network makes it possible for embeddings aggregate and update between vertices (detections and trajectories). We both introduce priori confidence to detection attention and trajectory attention, which consider the interaction between occluded objects in the same frame. Based on MHT framework, we train two graph networks for clustering in adjacent frame and association between long spaced tracklets. Experiments on MOT17/20 benchmarks demonstrate the significant improving in tracking accuracy of proposed method and show state-of-the-art performance for MOT with public detections.