Video Motion Detection Based On Optical Flow Field

This thesis is focused on the technique of the motion object detection based on optical flow in video sequences. Motion detection is an extremely important problem in the field of computer vision. Video images are composed of a series of frames, which have such as rich data, strong correlations between nearly images and dynamically changed mode in time domain, these characters make motion detection, segmentation and recognition becoming possible. Motion detection has many important applications not only in industry and military affairs, but also in docks, public, banks, vehicle surveillance system, and so on.The process of motion detection based on optical flow includes image preprocessing, optical flow computation, the utilizing optical flow computation, motion object extraction. The main works are as follows:1. Because of some exterior or interior influence, the images are usually polluted by noise. In order to improve the optical flow accuracy, the images should be smoothed before we analyze them. In this paper, we introduce some classical denoising methods. We adopt average filter in realization to improve the computation speed.2. The purpose of motion detection is getting the object from the complicated background. The optical flow methods have been put into a better application in motion detection. It's difficult to obtain the satisfactory results because of the influence of the dynamic environment, such as lighting changes, the shadow of motion object, swaying branches, rippling water. It has faults as following: heavy computation burden, weak anti-noise ability, the weak stability and this method can not track the moving objects with high speed. In this thesis, some classical methods are introduced. Based on linear brightness model, we propose a robust estimation of optical flow method. A general linear brightness model is embedded in the gradient constraint equation, which improves algorithmic stability under large varying illumination. The experimental results show that the algorithm is robust for local and global illumination variation. According to these improvements, the differential optical flow algorithm based on multi-scale-space focusing can not only adapt itself to larger motion velocity and larger velocity change, but also alleviate the error propagation effect through adopting bilinear interpolation operator.3. Motion object is obtained by normalized optical flow vector data, and thenwe realize morphological filtering and region connecting of binary image.Through horizontal projection and vertical projection, the motion object isenclosed by rectangle, which is prepared for the motion tracking.The paper is organized as following five parts. Section 1 gives the introduction ofthe background, significance and development of the motion detection. Section 2introduces some classical methods in image preprocessing. Section 3 introduces someclassical methods. Section 4 expounds in detail the new algorithm and the experimentalresults. Section 5, motion object is extracted by region connecting, then the analysis ofmoving object and some experimental results on real images are given. At last, thepaper ends with some further research plans mentioned in Section 6.