Online tracking of ants based on deep association metrics: method, dataset and evaluation. (July 2020)

• Record Type: Journal Article
• Authors: Cao, Xiaoyan; Guo, Shihui; Lin, Juncong; Zhang, Wenshu; Liao, Minghong
• Is Part Of: Pattern recognition. Volume 103(2020:Jul.)
• Publication Date: 2020-07
• Subjects: Ant tracking -- ResNet model -- Mahalanobis distance -- Appearance descriptors; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107233
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We introduce an online MOT framework to track ant individuals. This framework combines both motion and appearance matching, which effectively prevents trajectory fragments and ID switches from long-term occlusion caused by frequent interactions of ants, achieving efficient and high-precision tracking. We obtain ant appearance features based on the ResNet model with cosine similarity metric, to track unlabeled ants for a long time in a fixed position camera. The experiments show that our method is successful and robust with only a small size (N = 50) of the training dataset, which makes it feasible to be applied in real applications with no need to construct a large training dataset. We construct a dataset of ant tracking with a total size of 46091 samples. We built the dataset following the standard MOT formulation. In contrast to an extensive collection of human tracking datasets, there are few datasets of ant tracking which are publicly accessible. We believe this dataset will benefit future works with relevant research objectives. Abstract: Tracking movement of insects in a social group (such as ants) is challenging, because the individuals are not only similar in appearance but also likely to perform intensive body contact and sudden movement adjustment (start/stop, direction changes). To address this challenge, we introduce an online multi-object tracking framework that combines both the motion and appearance information of ants. We obtain the appearance descriptors by using the ResNet model for offline training on a small (N=50) sample dataset. For online association, a cosine similarity metric computes the matching degree between historical appearance sequences of the trajectory and the current detection. We validate our method in both indoor (lab setup) and outdoor video sequences. The results show that our model obtains 99.3% ± 0.5% MOTA and 91.9% ± 2.1% MOTP across 24, 050 testing samples in five indoor sequences, with real-time tracking performance. In an outdoor sequence, we achieve 99.3% MOTA and 92.9% MOTP across 22, 041 testing samples. The datasets and code are made publicly available for future research in relevant domains.