The Moving Target Detection Based On Frame Difference

As one of the basic research in the field of computer vision, moving target detection hasbeen widely used in military and civil application.It has good prospects of development.Because of the instability of the external environment and the complexity of movements,there are many problems with the moving terget detection. Traditional detection methods havetheir own advantages, but good results generally can be obtained only on certain occasionsand conditions. After the research of the existing algorithms, research work is expanded fortwo aspects based on frame difference in the paper.Firstly, the method of camera motion estimation based on feature matching of SIFt isproposed for dynamic background. The local feature extraction of video adjacent frames isdone by SIFT algorithm. After the matching of feature points, the set of matching points isobtained. The matching result is filtered by gray statistics data of feature point neighborhood.Combining the six-parameter model and least square method, the motion parameters arecalculated. At last, the results of direct frame difference and the frame difference after motionestimation are compared in the experiment. It shows that the proposed method can do thecamera motion estimation accurately and remove the adverse effects on local moving targetdetection caused by background movement.Secondly, taking the accuracy and real-time performance as starting point, we propose amoving object detection method based on Mean Shift clustering. The method takes movingregions of binarized difference image as blocks and extracts the bounding boxes of them.Then the bounding boxes are clustered by Mean Shift algorithm. Considering the location andsize of bounding boxes, a decision function is defined within the class.The function is used todivide the results of initial clusters. At last, we get the full region of every moving target.Experiments prove that the proposed method can accomplish the complete detection ofmoving targets effectively and has good clustering performence. The method is simple anddoes not need much pre-processing work in the early phase. Also its anti-interference is strongand has good accuracy and real-time.