Occlusion-related graph convolutional neural network for multi-object tracking

Abstract

Multi-Object Tracking (MOT) has recently been improved by Graph Convolutional Neural Networks (GCNNs) for its good performance in characterizing interactive features. However, GCNNs prefer assigning smaller proportions to node features if a node has more neighbors, presenting challenges in distinguishing objects with similar neighbors which is common in dense scenes. This paper designs an Occlusion-Related GCNN (OR-GCNN) based on which an interactive similarity module is further built. Specifically, the interactive similarity module first uses learnable weights to calculate the edge weights between tracklets and detection objects, which balances the appearance cosine similarity and Intersection over Union (IoU). Then, the module determines the proportion of node features with the help of an occlusion weight comes from a MultiLayer Perceptron (MLP). These occlusion weights, the edge weights, and the node features are then served to our OR-GCNN to obtain interactive features. Finally, by integrating interactive similarity into a common MOT framework, such as BoT-SORT, one gets a tracker that efficiently alleviates the issues in dense MOT task. The experimental results on MOT16 and MOT17 benchmarks show that our model achieves the MOTA of 80.6 and 81.1 and HOTA of 65.3 and 65.1 on MOT16 and MOT17, respectively, which outperforms the state-of-the-art trackers, including ByteTrack, BoT-SORT, GCNNMatch, GNMOT, and GSM.