Research On The Detection Methods Of Moving Objects In Non-planar Scenes

The research on the detection of moving targets in non-planar is now widely usedin military reconnaissance, precise guidance, civilian video surveillance and so on. Formoving carrier camera systems, the moving and random vibration of the camera willintroduce interference to the detection of moving target. When reconnoiter thenon-planar scenes closely, with respect to shooting distance, the depth of the scene isnot negligible, which will produce parallax. The parallax will be detected as movingtargets, which influences the movement detection. The following research are done toremove the moving and shaking of the carrier, and the interference of non-planarscenes.First of all, the movement of the carrier will change the background, it’s hard toseparate background motion and target motion. A feature-based image registrationmethod is proposed to compensate for camera movement. Three methods for featurematching: Harris corner, SIFT descriptors and SURF descriptor are used for theregistration. Make a comparative analysis of the above-mentioned feature-basedregistration algorithms from the computational complexity and registration accuracy toget a suitable image registration algorithm.Secondly, the carrier in the sky will shake inevitably, making the backgroundinstable and impacting moving detection. Taking the problem into account, digitalimage stabilization methods is used to eliminate high-frequency jitter. The researchfocused on the image stabilization method based on gray projection. In motionestimation based on gray projection method, the distribution of gray is affected bymoving target, bring error into motion estimation. Gray projection on parts of thepicture is produced to eliminate the impact of moving target. For the gray projectionmethod can only estimate the translational motion, circular gray projection method isadopted, transforming the video from Cartesian coordinates to polar coordinates toestimate the rotation.Furthermore, the gradient suppression method is firstly used to filter out parallax,which is produced by non-planar scenes. In the case of small-angle rotating this methodcan remove the parallax effectively. The parallax pixels can not be removed completelyin the case of large rotation. A method combining gradient suppression and epipolarconstraint is proposed to deal with the problem. The remaining pixels after gradientsuppression are applied to the epipolar constraint, so that all the independent movingpixels are detected.Finally, gray projection method is proposed to classify the dispersive movingpixels, then the boundary of independent moving target is obtained. The accurate detection and segmentation of the moving targets are achieved through simulationexperiments, proving the effectiveness of the program.