Background Modeling And Application Based On Model-to-model Distance In Complex Scenes

Background modeling has important research significance since it's an important task of object detection and scene understanding.This paper focuses on background modeling using the Model-to-Model paradigm in complex scenes,and the main research of this paper is as follows:Firstly,we propose a novel background modeling approach using deep neural network.The traditional background modeling approaches do not make full use of the spatial and temporal information of pixels.In our deep neural network,three atrous convolution with different dilate are used in three branches to extract spatial information of the pixel with different neighborhoods,which break the limitation that extracting spatial information of the pixel from fixed pixel neighborhood.Specially,we sample multiple frame images from original sequential images with variable interval to capture temporal information,which help our background modeling utilize enough temporal information without unnecessary calculation.Compared with the classical background modeling approaches,our approach has good adaptability in different complex scenes,especially in the Change Detection Benchmark Dataset database Office scene F-Measure to 92.36%.Secondly,we extract features from images using deep neural network because that deep neural network optimizes the feature extraction method by learning,so that features have a better representation of key information of images.Then we use features to build feature model,which can describe pixels from different levels for feature model includes high layers fine information and low layers coarse information.And we initialize background model simultaneously.To measure the similarity between the pixel and its background model,we propose a Model-to-Model(M2M)paradigm.Lastly,we use minimum M2 M distance to decide whether a pixel belongs to background.Moreover,background of the pixel is updated using minimum M2 M distance,while background of its neighbor pixels using maximum M2 M distance.Experiments results show that our M2 M approach has better performance than state-of-the-art approaches in complex scenes.Compared to the background modeling approach using deep neural network,M2 M has better performace,especially in the Change Detection Benchmark Dataset database,the F-Measure in the Office scene is 95.94%.