Research On Vision-Based Object Detection And Tracking Algorithm

As an extremely important content in the field of computer vision,target detection and tracking not only brings convenience and security to us in our daily life,but also has great application prospects in the fields of intelligent medical and military parposes.The corresponding research has been valued by many scholars and institutions at home and abroad.This dissertation focuses on some problems in the process of target detection and tracking,such as light changes,background disturbances,scaling and target occlusion,and improve and optimize traditional algorithms.The main works and contributions are as follows:1.ViBe algorithm with multi-information fusion: In this dissertation,we aim to solve the problem that the detection result is greatly affected by the artificial threshold during the detection of the target by three-frame differential method.The experiment result shows that the improvement improves the accuracy of detection and makes the extraction of target contour more complete.In order to solve the problem that the ViBe algorithm uses image gray features only when background modeling,resulting in a low detection accuracy in background scenes with similar colors.Replace the gray value with V component in the HSV color space,which improve the color accuracy in background scenes with similar colors without increasing the amount of algorithm calculations.Inspired by SOBS,the Euclidean distance in the background model update strategy in ViBe is replaced with the distance formula in SOBS,which improved detection accuracy.After ViBe fuses color features,it fuses with the target contour(texture feature)obtained in the previous step,then process the image by morphology.Experiment results show that the improved algorithm has a good detection effect on targets,and improves the hole phenomenon greatly.2.Adaptive window scaling: Since KCF needs to select the target area manually in the first frame,we use the improved target detection algorithm in this dissertation to replace the manual frame to achieve the target self-detection function in the tracking process.At the same time,its tracking box size is fixed,and the idea of scale pyramid is applied.Construct pyramids of different scales for the target area,and determine the size of the tracking box in the current frame according to pyramid of the appropriate scale,which improves the speed of algorithm.By making the tracking box size adaptive to target,the tracking accuracy is improved.3.KCF algorithm based on SIFT feature extraction: Target tracking by KCF becomes worse when occlusion occurs,it may even fail if occlusion continues.Aiming at this condition,this dissertation introduces the fusion algorithm.Firstly,we need to detect whether the target is occluded in the current frame and once occlusion is detected,constructing a fusion feature model of the target area of the current frame.And perform model matching operations with subsequent frames,until the target area is detected again.Experiment results indicate that the improved algorithm has better tracking effect when the target is occluded.And it can still be successfully tracked when the target recovers from occlusion to an integrated target.