3D VISUAL TRACKING USING A SINGLE CAMERA

automated surveillance and motion based recognition. 3D tracking address the localization of moving target is the 3D space. Therefore, 3D tracking requires 3D measurement of the moving object which cannot be obtained from 2D cameras. Existing 3D tracking systems use multiple cameras for computing the depth of field and it is only used in research laboratories. Millions of surveillance cameras are installed worldwide and all of them capture 2D images. Therefore, 3D tracking cannot be performed with these cameras unless multiple cameras are installed at each location in order to compute the depth. This means installing millions of new cameras which is not a feasible solution.
This work introduces a novel depth estimation method from a single 2D image using triangulation. This method computes the absolute depth of field for any object in the scene with high accuracy and short computational time. The developed method is used for performing 3D visual tracking using a single camera by providing the depth of field and ground coordinates of the moving object for each frame accurately and efficiently. Therefore, this technique can help in transforming existing 2D tracking and 2D video analytics into 3D without incurring additional costs. This makes video surveillance more efficient and increases its usage in human life.
The proposed methodology uses background subtraction process for detecting a moving object in the image. Then, the newly developed depth estimation method is used for computing the 3D measurement of the moving target. Finally, the unscented Kalman filter is used for tracking the moving object given the 3D measurement obtained by the triangulation method. This system has been test and validated using several video sequences and it shows good performance in term of accuracy and computational complexity.