Name
Abstract | ||
|---|---|---|
Noisy detection and similar appearance lead to deteriorated mis-identification and id-switch in Multi-Object Tracking (MOT). To address these problems, we propose a novel Nearest Optimal Template Library (NOTL) associated with two tailor-made methods based on the NOTL. Here, the NOTL is a historical sample set of the tracked objects, and the elements in the NOTL are closest to the complete object at the current instant. It provides reliable appearance information of the object. Then, we use the single object tracker (SOT) for position prediction, and spatio-temporal network for appearance modeling. They can alleviate mis-identification and id-switch problems, respectively. Besides, the triplet loss is used to train our spatio-temporal network further improves the performance. The proposed algorithm achieves 55.3% and 55.1% in MOTA on challenging MOT16 and MOT17 benchmark datasets respectively. These results show our method is competitive with the previous state-of-the-art approaches. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1007/978-3-030-86362-3_27 | ARTIFICIAL NEURAL NETWORKS AND MACHINE LEARNING - ICANN 2021, PT I |
Keywords | DocType | Volume |
Multi-object tracking, Template library, Single object tracker, Spatio-temporal network | Conference | 12891 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 4 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Ran Tian | 1 | 0 | 0.34 |
| Xiang Zhang | 2 | 195 | 34.67 |
| Donghang Chen | 3 | 0 | 0.68 |
| Yujie Hu | 4 | 0 | 0.34 |