Research On Moving Target Detection,Location And Tracking System Based On Optical Flow Information

Moving target detection,location,and tracking are the core research topics in the computer field.They are with very significant value of research in video surveillance,military confrontation,intelligent transportation,etc.There are many relevant algorithms for moving target detection,location,and tracking.However,as new problems continue to emerge,there are still lots of theoretical and application difficulties that can be further researched and solved.This thesis takes the moving target in 3D space as the research object,takes the monocular camera as the research tool,and takes optical flow as the research starting point.Furthermore,integrates the methods of image processing,deep learning,clustering,and nonlinear filtering to propose a complete system incorporating moving target detection,spatial location,and stereo tracks.The main work of this thesis is as follows:In terms of moving target detection,to meet the requirements of large-displacement moving target detection under a large amount of noise interference and continuously improve the detection effect,this thesis carries out the research work in the first and second stages,respectively,based on traditional methods and deep learning strategies.In the first stage,the optical flow prediction and moving target extraction network based on conventional computer vision are constructed,including preprocessing images and improving the Lucas-Kanade optical flow method to calculate optical flow.Furthermore,a two-level moving target extraction strategy based on the combination of coarse and fine is proposed.In the second stage,based on the first stage,the optical flow prediction and moving target extraction network based on the learning algorithm are constructed to improve further the detection algorithm.Specifically,feature regions are extracted by enriching image preprocessing,applying improved PWC-Net computing optical flow,innovating K-means and aggregation clustering algorithms,and fusing the frame difference method.Experiments show that the proposed detection algorithm can accurately and ultimately obtain the moving target.In the aspect of moving target location,to obtain the 3D position of the moving target by only using a single frame image of the monocular camera,a monocular vision spatial location algorithm based on pinhole imaging is proposed according to the detection results.Firstly,the positioning model is established and combined with the camera parameters.Next,the coordinates of a single frame image are transformed to obtain the spatial orientation information of the target.Then,the target’s spatial coordinates and depth information are obtained by using the spatial orientation information,the camera’s installation position,and the geometric relationship.Experiments show that the location algorithm in this thesis is simple,economical,practical,and accurate,and the average relative error of location information is only 3.91%.In the aspect of moving target tracking estimation,to solve the 3D tracking of moving target through 2D image,a stereo tracking strategy based on Cubature Kalman Filter is proposed according to the positioning results.Firstly,the motion models of moving targets are established.Moreover,the system state space expression is constructed with the speed and position of the moving target as state variables and the pixel coordinates of the moving target as the measurement output.Finally,the different models are fused with the Cubature Kalman Filter to realize the stable tracking of the moving target’s spatial position and velocity.In addition,the Extended Kalman Filter is introduced in the experiment to verify the feasibility,advanced,and stability of the Cubature Kalman Filter.Experiments show that the mean value of RMSE,Manhattan distance,and Euclidean distance of Cubature Kalman Filter tracking strategy concerning 6D tracking variables are significantly better than Extended Kalman Filter tracking strategy in motion prediction time.