Tracking Multiple Indistinguishable and Deformable Objects Based on Multi-Anchor Flow With Annular Sector Model

Abstract

Most of current multi-object tracking (MOT) methods solve the task of identity assignment mainly by using the distinguishable features and predictable motions of nearly rigid objects. However, when the objects are indistinguishable, non-rigidly deformable and erratically motorial, it is challenging to differentiate the identities and regions of multiple objects in occlusions. In this paper, a novel method based on multi-anchor flow with annular sector model (ASM) is proposed to handle occlusions and complex motions for tracking multiple of these kinds of objects. The multi-anchor flow is built on anchor points which are extracted from the union contours of deformable objects. By optimizing the multi-anchor flow over frames, the proposed method solves the task of identity assignment when multiple objects are in clustered. Meanwhile, the annular sector model is also combined into the tracking method not only to rectify the results of identity assignment, but also to depict the poses or regions of deformable objects in occlusions. In our experiments, the proposed method is evaluated on two publicly challenging dataset for tracking multiple Drosophila Larvae. The results demonstrate that the proposed method has better performance on both of the accuracy of identity assignment and the computational time when compared to other multi-object tracking algorithms.

Keywords

• Multi-object tracking
• larvae tracking
• annular sector model
• multi-anchor flow
• indistinguishable objects tracking