A Fast 3D Multi-Object Tracking Method Based On Motion and Appearance Features

Abstract

3D-MOT(3D multi-target tracking) is an important direction in the research of autonomous driving and intelligent robots, which can provide a rich and reliable dynamic representation of environment for modules such as path planning. This paper proposes a 3D-MOT method that fuses 3D and 2D input data. This method takes into account the motion characteristics and appearance characteristics of the target during the data association process, and uses the Kalman filter to predict the target in 3D and 2D space respectively. State, and proposed a target motion similarity measurement method that integrates motion features in different spaces; at the same time, this method also designed a network for extracting the apparent features of the detected target, and based on this, proposed a similarity feature Finally, an integrated similarity calculation method integrating motion similarity and apparent similarity is proposed for data association. On the basis of ensuring various tracking indicators as much as possible, the purpose of 3D multitarget tracking is achieved with a small amount of calculation, which can meet the real-time algorithm requirements of service robots and other platforms with low detection accuracy and limited computing power.