Person 3D Tracking and Action Recognition through Stereo Camera

Abstract

The use of depth information to estimate the precise position of an object in the world coordinate system has attracted the attention of researchers in 3D multiple object tracking. Recent studies have obtained depth information from expensive LiDAR systems to achieve accurate results in vehicle tracking tasks. Some studies have used pixel information from a single image obtained from a monocular camera to predict depth, which is cheap but has significant limitations in accuracy. To accurately track small objects like the person and analyze their behaviors with a lower cost, in this paper we use stereo camera and propose a combination of 2D detection and 3D tracking method to implement a 3D tracking analysis system. Our approach is less costly and the system can be used conveniently. Our code is publicly available at https://gitlab.com/syo093c/fuminsyoStereo.