Research On Moving Object Detection In Complex Scenarios

Moving object detection is to intelligently detect and segment moving objects that are interested by observers from video data,and has become one of the fundamental and core technologies of smart video analysis,because it can facilitate further analysis and understanding of those moving objects.This gives rise to its increasingly wide application to smart video surveillance,intelligent traffic management,human-computer interaction,self-driving cars,etc.Aiming at resolving challenges in real applications and difficulties in theoretical research of the moving object detection,in this dissertation,moving object detection methods for complex application scenarios are deeply studied.Main research work and achievements of this dissertation are listed as follows.(1)To improve the ability of distinguishing tiny differences of features between foreground objects and background,a moving object detection algorithm based on human visual brightness sensitivity using pixel-based features for background modeling is proposed.Enlightened by the perceptual characteristics in brightness discrimination of mature human visual system(HVS)under different background illumination conditions,we first deduce a Weber ratio suitable for complex images as background environments,and then use this Weber ratio to set adaptive matching thresholds for background samples of different brightness in the reference background model.In this way,the capability of perceiving subtle differences between the foreground objects and the background is effectively enhanced,and the accuracy of the foreground detection is greatly improved.Experimental results demonstrate that,the proposed algorithm has outstanding ability to deal with the camouflage challenge.(2)To adapt to dynamic changes and represent accurate background of the scenes,two moving object detection algorithms using complex features for background modeling are proposed.In order to cope with dramatic changes in complex scenarios such as intense disturbance from dynamic backgrounds,inspired by global properties in visual perception of the HVS,through digging complex spatial features of scenes,a moving object detection algorithm utilizing complex features that contain global information of scenes for background modeling is proposed.This method modifies traditional pulse coupled neural network(PCNN)to simulate the perceptual process of color scenes by the HVS?and enhances expressive force of the PCNN in the overall characteristics of scenes.Response features obtained from the modified PCNN,containing global information of scenes,are employed to build the background model,which will substantially strengthen adaptability of the background model in dealing with complex and volatile environments.Experimental results show that,the proposed algorithm is highly robust to background perturbance such as dynamic backgrounds.In order to comprehensively take advantage of multisource video images to obtain a more accurate background description of the scenes,a moving object detection algorithm based on fusion model of multisource features is proposed,by making use of complex spectral features of scenes from both visible and thermal videos.According to the joint background model that fuses multisource spectral information,the proposed algorithm implements joint matching and decision-making of multisource features extracted from the current input data,which greatly improves its ability to fuse and utilize the complementary information from both visible and thermal videos.Experimental results prove that,the proposed algorithm not only avoids the impacts of illumination changes,shadows and camouflaged foreground objects in visible videos,but also overcomes the issues of halos and thermal reflections in thermal videos.(3)Considering the lack of adaptability to actual environmental conditions when existing subspace learning-based moving object detection methods represent the background,we introduce prior information,and then propose a moving object detection algorithm based on online learning of background low-dimensional subspace and a moving object detection algorithm based on conditional separation of background independent subspace.To make the learning of background low-dimensional subspace adapt to complex changes of scenes,online robust principal component analysis(ORPCA)is improved by setting adaptive weights for the sparse term,and then employed to dynamically estimate and learn the background low-dimensional subspace in an online manner.The algorithm utilizes a feedback mechanism to adaptively adjust the degree of flexibility of subspace decomposition to sparsity of the foreground objects,which will effectively enhance its adaptability to dynamic backgrounds and various sparsity of the foreground objects.Experimental results demonstrate that,the proposed algorithm effectively suppresses the impacts of camera jitter to moving object detection without any anti-jitter processing,and is highly robust to complex changes of real scenarios.To comply with the practical situation in which an observed video is actually a nonlinear mixture of the foreground objects and the background,we transform moving object detection into a nonlinear blind source separation(NBSS)problem,and introduce the nonlinear independent component analysis to implement the NBSS from input video data,and further use prior information about the background as constraints.This makes the accuracy of independent background subspaces estimation and foreground detection considerably improved.Experimental results show that,compared to linear separation manner,nonlinear separation manner which is more consistent with the actual situations attains more accurate detection results,in particular,the proposed algorithm can effectively solve the problem of intermittent object motion.(4)Most existing methods require a separate training phase to learn to separate the foreground objects and the background,and may have low detection accuracy resulted from factors such as insufficient learning.To solve this problem,a moving object detection algorithm based on three-dimensional(3D)discrete wavelet transform(DWT)is proposed.The proposed algorithm has no learning and training process,but takes advantage of the multiscale analysis feature of the 3D DWT,and separates the foreground objects and the background(possessing different temporal characteristics)directly into different 3D wavelet subbands.After this,the subbands corresponding to the background are directly zeroed,and disturbance coming from non-foreground objects(such as illumination changes,noise,etc.)in the other subbands are suppressed.In this way,foreground objects can be segmented accurately and rapidly.Experimental results prove that,the proposed algorithm has great potential in conquering the tough challenge in many real applications—situations that lack training and learning opportunities.