Real-time Tracking of Multiple Objects by Linear Motion and Repulsive Motion

Real-time Tracking of Multiple Objects by Linear Motion and Repulsive Motion

Abstract: Successful multi-object tracking requires consistently maintaining object identities and real-time performance. This task becomes more challenging when objects are indistinguishable from one another. This paper presents a Bayesian framework for maintaining the identities of multiple objects. Our semi-independent joint motion model (SIMM) solves the coalescence and identity switching problem in real time. This joint motion model is a non-parametric mixture model that simultaneously captures linear motion and repulsive motion. Linear motion is a constant velocity model, while repulsive motion is described by a repulsive potential in MRF. By maintaining multimodality from multiple motion models, we can infer the appropriate motion model using image evidence and consequently avoid many identity switching errors. Moreover, we develop a new sampling method that does not suffer from the curse of dimensionality because of the availability of high-quality samples. Experimental results show that our approach can track numerous objects in real time and maintain identities under difficult situations.

Author: Lejun Shen(沈乐君), Zhisheng You(游志胜) and Qing Liu(刘青)

Download: PDF (全文) , 演示程序 源代码

关键词：实时跟踪, 实时多目标跟踪, 身份交换错误, 粒子滤波器, MRF, 多目标跟踪, 视觉跟踪,  马尔可夫随机场, 自助法, 重要性采样