Moving Object Detection Technology Study Based On Fractal Analysis And Optical Flow Estimation

Moving object detection is a key technology in optical imaging system and ahot research topic in computer vison field. Noval theories and ideas arecontinuously introduced to this topic in recent years, which forcefully promote thetechnology progress, bringing in a lot of new image processing methods formoving object detection. However, there still exist a few unsettled problemswhich influence the performance improvement of the object detection systems.It is necessary to make a further study on the topic.The paper focuses on the moving object detection in images with complexbackgrounds, two methods which are applicable for static background anddynamic background are respectively developped. The first method’s detectionprocess relies on a preprocessing precedure which estimates the motion state ofevery image area by its fractal dimensions. The latter one describes the varietyof image intensity between two frames using statistical fractal model anddetermines optical flow field, then implements detection process based on theresults of flow field calculation. The main contributions in the paper are asfollows.1) The moving object detection technique based on three frames jointscatter diagrams (3F-JSD) is proposed, which solves traditionalproblems such as occlusion-unocclusion, illumination-variety andnoise-disturbance. The technique judges the motion state of an imagearea according to the structure characteristic of its3F-JSD, whichdifferentiates moving areas from still areas by sorting their3F-JSD, andextracts object from the moving areas by segmenting their3F-JSD. Testresults show that the technique can accurately extract moving objectsfrom static-background images.2) The basic optical flow equation based on the fractal Brownian motionmodel (BE-FBM) is deduced to determine the flow of natural images.We firstly argued that the intensity fields of natural images can be welldescribed using the FBS model, which was proved by both theoreticaland experimental methods; then proposed that the FBM model shouldbe used to describe the time-domain intensity fields of natural imagesequences; finally deduced the BE-FBM based on FBM model. Testresults show that the BE-FBM is more applicable than the traditionalbasic equation based on the hypothesis of intensity constance (BE-HIC)for determining flow field of natural scenes. 3) The local flow technique using the Least-weighted-median-of-squaresmethod (LWMS) is proposed, which can figure out correct flow values incorner areas. The technique’s first step is picking up reliable pixels fromlocal widow using the score strategy; then the initial reliable pixels areadopted to determine initial flow value, which is corrected using theneighbor-correcting method later; after that, flow remainders arecalculated using the initial flow value, according to which final reliablepixels are filtrated; ultimately the final flow value is computed byadopting the remainder reliable pixels. Pixel-filtration andneighbor-correcting enable the LWMS method to estimate the flow inthe image’s coner areas in high accuracy.4) A multi-technique fusion (MTF) method simultaneously implementingocclusion detection, flow diffusion and corners matching is proposed,which can figure out correct flow fields for mismatched areas and smallobjects when the displacements between neighbor image frames arelarge. The MTF method firstly constructs the cost function using basicequation constraint, mismatching constraint, anisotropic smoothnessconstraint and flow constraint deduced from the matching method; thenthe optical flow field is determined using Lorentz method under themulti-scale frame. The multi-scale technique together with the matchingmethod can effectively figure out correct flow values for small objects,while the mismatching constraint together with the anisotropicsmoothness constraint can effectively reckon out correct flow values formismatching areas.5) The multiscale normalized cut (M-NCut) method is adopted to exactlysegment the flow field, which can be implemented through verticesmerging and vertices segmentation. In the merging process, the methodregards each flow value as a vertex and merges similar vertices level bylevel till salient vertex appears. In the segmentation process, the statevalues of the salient vertex are transferred level by level till the initialflow field, such that a segment is obtained. The merging process andsegmentation process are repeated by turns until the whole field issegmented.6) The diagram flow of moving object detection method based on opticalflow field (MOD-OFF) is presented, which is applicable for dynamicbackground. The method’s performance is tested with variety of typicaldynamic-background image sequences. The results show that theMOD-OFF method can perfectly extract objects in images with varaint characteristics, showing a performance far better than other methods.Meanwhile, the illumination-variety and noise-disturbance experimentsdemonstrate the method is robust.