Optical Flow Field Segmentation Algorithm For Moving Target Detection

Moving object detection in dynamic scenes is the key technology of automatic driving and cruise monitoring,improving the detection accuracy and detection efficiency has become the focus of current research.In this paper,a moving object detection method based on block flow field segmentation and dense optical flow field segmentation is proposed from the viewpoint of optical flow vector distribution.The block flow field segmentation algorithm divides the image into several fixed-size rectangular blocks,based on the local constraint of the optical flow field,the matching criterion is constructed with the brightness conservation constraint and the gradient conservation constraint.The similarity block is searched in the reference frame by small and large diamond template and the coordinate of the best matching block is used to calculate the block's optical flow vector.Through adaptive K-means clustering algorithm to achieve the optical flow field transform to the movement category,and finally optical flow field segmentation and color image segmentation are fused to improve target hole and rough edge.TV-L1 dense optical flow field segmentation algorithm is based on variational optical flow model.Building recombination data items by brightness conservation constraint and gradient conservation constraint,using global smooth model to create a priori item integrating total error by 1-norm.Through total variational method we can transform,optical flow filed calculation into solving a large non-linear equation set.Removing the non-linear character through dual fixed point iteration and getting the optical flow vector at each point by over-relaxation iterative method.In order to improve the efficiency of optical flow estimation,the Gaussian pyramid is embedded in numerical procedure to estimate optical field coarse-to-fine.Using self-adaptive K-means cluster method during optical field segmentation allowed us to extract moving object area in pixel-level.The comparative experiments and analysis illustrate that BMD algorithm and TV-L1D algorithm achieve better result than SUD.BMD algorithm realized high accuracy real-time detection in 33ms per frame.TV-L1D algorithm involves a large number of iteration and float calculation make it difficult to detect object in real-time.