The Research Of Detection And Tracking Methods Of Moving Objects In Dynamic Scenes

Moving object detection and tracking is a research hotspot in the field of computer vision,and it is also two continuous related issues.They are the technical basis of video behavior analysis and behavior recognition,and are widely used in military and civil fields.At present,a number of studies have been devoted to this task,but there are still some problems to be further studied.Aiming at the problems of real-time and accuracy in detection of moving object in dynamic scenes is difficult to take into account.,a fast multiple moving object detection framework is proposed,which combines several mainstream technologies.This method makes use of the principle that corresponding corners of the background in successive frames meet the constraints of epipolar geometry in video sequence,the basic idea of detection is divided into the following three steps: 1)for the two adjacent images,extracted Harris corners in the first frame image,then get corresponding points in the next frame of these corners using LK optical flow method based on Pyramid layer;2)taken these points pairs as the training set,a fundamental matrix satisfying the constraints of epipolar geometry is trained as a classifier to identify the foreground/background corners,the fundamental matrix is estimate iteratively using RANSAC algorithm;3)clustered corners of foreground obtained from the previous step,each class represented a moving object area.Aiming at the problems of affecting tracking accuracy caused by complex shooting environment and object motion in moving object tracking,a Mean Shift object tracking method based on Kalman filter is proposed.The main tracking process is as follows: 1)Kalman filter predicts object using motion state of object in time domain 2)Mean Shift algorithm regards the predicted position as the initial value to search locally with the adaptive object region size;3)using the threshold to compare with the maximum similarity function value,if the value is larger,the result of local search is the final object tracking region for this frame;if the value is smaller,Mean Shift algorithm regards the object center of the previous frame as the initial value to search globally with the adaptive object region size again,and the threshold is used again to compare with maximum similarity function value to determine the final tracking region is the global search result or the Kalman predicted result.Experimental results show that comparing with the Dual-Mode SGM model and extended background subtraction detection method,the proposed detection method not only simultaneously detect multiple moving objects in dynamic scenes,but also owns high detection accuracy and fast speed.Comparing with the kernel correlation filtering tracking method and the discriminative correlation filter with channel and spatial reliability tracking method,our tracking method can track moving objects in dynamic scenes with deformation,rotation,occlusion,sudden change of motion of objects and so on,and the tracking accuracy is high and the needs of real-time tracking can be met.The innovations of this paper are as follows: 1)A framework of moving object detection in dynamic scenes is proposed,which combines epipolar geometry,LK optical flow and other methods.The estimation method of fundamental matrix satisfying epipolar constraints is improved,and the estimation accuracy is improved.2)A real-time tracking method of moving object in dynamic scenes is proposed,which combines Mean Shift algorithm with Kalman filter.The algorithm can not only adjust the size of matching window adaptively,but also can determine whether global search is performed according to the threshold of matching after local search to reduce the computation.