Multi-object Tracking with Conditional Random Field

Abstract

The goal of online multi-object tracking (MOT) is to estimate the tracks of multiple objects instantly with each incoming frame using up-to-the-moment information. Conventional approaches to MOT do not account for motion of the camera or background. Also, traditional optical flow features or color features do not produce satisfactory discrimination. This paper attempts to deal with such problems by turning the online MOT problem into the problem of minimizing the global energy by constructing a conditional random field model. At the same time, structural information and discriminative deep appearance features are integrated together to make data association more accurate. In this way, individual objects are not only more precisely associated across frames, but are also dynamically constrained with each other in a global manner. We conduct experiments using the MOT challenge benchmark to verify the effectiveness of our method and achieve competitive results.