Multi-Object Tracking (MOT) which aims to estimate trajectories for objects of interest in videos, is a crucial component in intelligent transportation systems and has gained considerable attention from both academia and industry. However, existing MOT approaches apply simple motion and appearance features, ignoring the spatial-temporal correlations of interactions with other objects. To fill this gap, this paper proposes a novel MOT framework ST-Tracking to extract spatial-temporal features for tracking objects, which is trained in an end-to-end manner. Specifically, a directed graph is first constructed to represent the spatial-temporal correlations, where the node attributes include both geometric and appearance features, and the edges are deliberately defined to embrace location and topology cues. Then, a spatial-temporal graph convolutional module is proposed to aggregate information from both spatial and temporal dimensions, thus extracting discriminative features. Extensive experiments on the challenging KITTI benchmark demonstrate the effectiveness and superior of ST-Tracking as compared with several state-of-the-art methods.
ST-Tracking: Spatial-temporal Graph Convolution Neural Network for Multi-object Tracking
19.09.2021
1567949 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Object tracking in surveillance videos using spatial-temporal correlation graph model
British Library Online Contents | 2015
|Multi-Object Tracking Using Dynamical Graph Matching
British Library Conference Proceedings | 2001
|Multi-object tracking using dynamical graph matching
IEEE | 2001
|