Research On Moving Object Detection Based On Saliency

With the rapid development of microelectronics manufacturing process,digital encoding and network communication technology,videos have been widely integrated into and changed our lives.Although such videos form an intuitive,information-rich expression for spreading knowledge or sharing experience,it also poses a challenging research question,i.e.,how to efficiently and intelligently analyze video data? Intelligent video analysis covers object detection,target identification,behavior judgment,scene understanding and so on,in which moving object detection,as the basis and precondition of the latter processing,has attracted much attention over the years.Also there have accumulated a large number of research results on this issue.However,the practice indicates that moving object detection in the real-life scenarios is far from mature.With respect to the human visual system,even the most superior algorithms are eclipsed.Allowing computers to understand videos and process these information in the attractive notation which accords with our human mental data representation is still a great challenge to researchers.This paper focuses on moving object detection from visible image sequences based on saliency,inspired by the visual attention mechanism.The major contributions of this paper are:(1)We propose a moving object detection and segmentation algorithm based on temporal information.Firstly,the motion saliency map is calculated with the continuous symmetric difference of adjacent frames.In order to get more accurate and robust detection results,the proposed method greatly expands the size of the time window,which well pops out the moving objects and meanwhile inhibits the perturbations of the background.Then,based on the motion saliency map,the information entropy and fuzzy theory are utilized to adaptively extract the moving objects in the scene.The proposed method does not require any prior knowledge nor interactions by users.The detection results on a number of publicly available datasets demonstrate its effectiveness and advancement.(2)We introduce a motion saliency detection method based on gray difference.To be specific,the difference is modeled as two parts,i.e.,symmetric frame difference and background sample difference.After considering the robustness and complexity,we apply a nonlinear fusion scheme to build the final motion saliency map by combining the above two features.Experiments on a variety of visual surveillance databases verify that the proposed algorithm not only eliminates the phenomena of trailing and moving blur,but also suppresses the ghost in the detection results.In addition,with a simple adaptive thresholding method,we demonstrate how the generated saliency maps can be used to create high-quality segmentation masks for moving object extraction.(3)We design a deep convolutional neural network(DNN)to detect motion saliency at the condition of low illumination.Firstly,labeled data under different environments are used to train a deep network with 376,700 parameters,9 levels,of which the maximum width is 128.Taking into account the difference between adjacent frames directly reflects the motion cues in the scene,so other than the original image sequence,we introduce the corresponding frame difference maps as the training data.Unlike the traditional DNNs,we do not reduce the dimensionality of the input data during the feature extraction process,and keep the map size unchanged.The output of the network is the full resolution saliency map of the input image.Being trained with the largest artificial dataset,the detection results of the network under low-light conditions demonstrate its effectiveness.(4)We present a saliency detection method based on fractal characteristic.Firstly,the proposed method gets the saliency map with the difference of fractal characteristic between man-made objects and natural background.Then,a threshold is adaptively calculated using the maximum entropy method to binarize the saliency map and get the candidate area.Finally,we take the minimal circumscribed rectangle of the candidate area as the initial profile curve and apply the active contour model to make the evolving curve moving to the object boundary.The proposed method can effectively solve the problem of active contour model associated with curve initialization.Also,experimental results on different image sequences show that it can accurately detect and segment the man-made objects in motion and is robust to the perturbations of the background.The work mentioned above,as the underlying technology of moving object tracking,classification,identification and behavior analysis,can be widely used in the fields of intelligent video surveillance,automatic driving,battlefield environmental monitoring,space object detection and so on.