Real-Time Object Motion and 3D Localization from Geometry.

摘要

Knowing the position of an object in real-time has tremendous meaning. The most widely used and well-known positioning system is GPS (Global Positioning System), which is now used widely as invisible infrastructure. However, GPS is only available for outdoor uses. GPS signals are not available for most indoor scenarios. Although much research has focused on vision-based indoor positioning, it is still a challenging problem because of limitations in both the vision sensor itself and processing power. This dissertation focuses on real-time 3D positioning of a moving object using multiple static cameras. A real-time, multiple static camera system for object detection, tracking, and 3D positioning that is run on a single laptop computer was designed and implemented. The system successfully shows less than +/-5 mm in real-time 3D positioning accuracy at an update rate of 6 H z to 10 Hz in a room measuring 8x5x2.5 meters. Implementation and experimental analysis has demonstrated that this system can be used for real-time indoor object positioning.;In addition, 'collinearity condition equations of motion' were derived that represent the geometric relationship between 2D motions and 3D motion. From these equations, a 'tracking from geometry' method was developed that combines these collinearity condition equations of motion with an existing tracking method to simultaneously estimate 3D motion as well as 2D motions directly from the stereo camera system. A stereo camera system was built to test the proposed methods. Experiments with real-time image sequences showed that the proposed method provides accurate 3D motion results. The calculated 3D positions were compared with the results from an existing 2D tracking method that uses space intersection. The differences between results of the two methods were less than +/-0.01 mm in all X, Y, and Z directions. The advantage of the tracking from geometry method is that this method calculates 2D motions and 3D motion simultaneously, while other tracking methods need an additional triangulation step to estimate 3D positions.