Research On Moving Target Extraction Method Based On A Moving Camera

With the continuous advancement and development of software and hardware technologies,the emergence and application of mobile data acquisition devices such as drones and measuring vehicles,the promotion of new data acquisition devices such as GoPro and panoramic cameras,.The related algorithms in the field of computer vision originally designed for fixed camera are no longer suitable for new scenarios,that is,the algorithms for fixed cameras need to be expanded and improved to some extent,driven by real-life needs.The moving target recognition algorithm is often used in the field of video surveillance.Since the data acquisition device for video surveillance is usually a fixed camera,common moving target recognition algorithms are often used for fixed cameras,which is not suitable for the new generation of monitoring systems combining PTZ,drone and other technologies.Therefore,combined with the characteristics of the new generation of data acquisition platform,it is of great practical significance to extend the motion target recognition algorithm limited to fixed cameras to the related research work in the field of motion data acquisition platform.In view of the shortcomings of the existing background difference algorithms in the field of moving target detection,this paper combines the characteristics of data acquisition platform and other existing technologies,discusses and implements the background difference algorithm suitable for motion data acquisition platform.The main results are as follows :(1)An algorithm for estimating camera motion by multiple homography transformation is proposed.In this paper,the transformation matrix of images and common motion estimation algorithms are studied.Combined with feature point matching,homography transformation and RANSAC,a camera motion estimation algorithm based on multiple homography transformation is proposed.The algorithm uses ORB feature point matching and cyclic RANSAC.The multi-uniform transformation matrix of adjacent frame images is calculated as the camera motion estimation result.The multi-uniform transformation results represent the pixel mapping relationship in different background planes,and the global optical flow compensation is used to eliminate the moving targets.Experiments show that the algorithm can obtain suitable multi-uniform transformation on the basis of removing the feature points of moving targets.(2)An improved Vibe algorithm based on multiple homography transformation is proposed.In this paper,the common background difference algorithm is studied,and the more popular Vibe algorithm is studied in detail.According to the shortcomings and characteristics of Vibe algorithm,combined with multiple homography transformation to estimate the camera motion algorithm,an improved Vibe based on multiple homography transformation is proposed.The algorithm discusses the different mapping relationships of adjacent frame pixels caused by camera motion and gives corresponding solutions for various situations.It improves the background model update and foreground decision part of the original Vibe algorithm.Experiments show that the improved algorithm overcomes the shortcomings of the ghost region and can greatly suppress the situation that the camera is misdetected as foreground due to camera movement,and outperforms MBS,FTSG,SuBSENSE and CwisarDH in accuracy,specificity and false alarm rate.(3)A shadow removal algorithm based on HSV sample set strategy is proposed.In this paper,the characteristics of shadow regions and the common techniques of shadow region detection are studied.Combined with sample set strategy and HSV color space,an improved Vibe algorithm based on HSV sample set strategy is proposed,and HSV sample set strategy is integrated into multi-singularity-based transformation.Experiments show that the improved Vibe algorithm based on HSV sample set strategy has obvious shadow removal effect on static camera dataset,and it is obviously superior to FTSG,SuBSENSE and CwisarDH in shadow false alarm rate.In the dynamic camera dataset,the HSV sample set strategy still has some shadow removal effects on the improved Vibe algorithm which is based on multiple homography transformation.The algorithm in this paper can achieve 20 FPS processing efficiency when the resolution is 320*180.