Research On Scene Segmentation Under Sudden Illumination Changes By Feature Fusion

Scene segmentation technique is the core technology in the field of objectdetection, which has great academic and practical significance. Although differentmethods on how to model scenes had been proposed, they would still be hard to get theexact scene segmentation results when the illumination changes rapidly. To solve thisproblem, this paper concentrates on the scene segmentation method for the changingillumination. The main study can be summarized as follows:①In this paper, the state of art of the video surveillance at home and abroad wassurveyed. At the same time, we compared and analyzed the existing intelligent videosurveillance technologies. The main challenges in scenario modeling technology weredescribed, which led the coming chapters.②We described the basic theory of moving object detecting, in which theintroduction of the scenario modeling techniques was emphasized. The state of art ofvideo surveillance scenario modeling were described, and their advantages anddisadvantages were compared. Meanwhile, as the light’s sudden change would have ahuge disturbance in the scene models, there were many deficiencies in the classicalmethods. We proposed a novel Gaussian mixture modeling method based on featurefusion to reduce the impact of the light change on the scene segmentation.③We performed research on three kinds of illumination invariant features,described Texture, ZNCC and contour features, and introduced the conrrespondingfeature extraction methods. Next, we verified such illumination invariant features’robustness and the possibility of fusion by experiment. Meanwhile, for the currentcontour feature extraction problems, we proposed a new process to make the extractedcontour more stable, smooth, continuous, and more suitable for fusion.④We built a scene model through integrating the global illumination function intothe framework of Gaussian mixture models. Then, the illumination invariant features,which were extracted in③，were combined for scene segmentation in two steps:Specifically, the ZNCC and textures were combined in the first step, and the contourwas integrated in the second step. Finally, the experimental results showed that ourmethod could effectively improve the accuracy and robustness of the foregroundsegmentation.