Research On Target Detection And Tracking Algorithms Based On Omnidirectional Vision Sensor Node

In recent years,with the development of computer vision technology and wireless multimedia sensor network,security monitoring based on vision sensors and combined with certain algorithms becomes more intelligent,which make it become one of the hot issues in this subject.The omnidirectional vision sensor node composed of multiple fisheye cameras is an important component of the WMSN.It is generally independently laid in the unattended environment to complete assigned tasks.On the background of omnidirectional vision sensor node in WMSN,the target matching and tracking in multiple cameras with overlapping field of view is studied.The main contents of this thesis are as follows:Due to the application of the fisheye camera,the perspective camera model and relevant basic knowledge are studied.At the same time,the fish-eye distortion and corresponding model are introduced.An improved automatically fish-eye camera calibration method is proposed based on division model.The curves in the image are extracted and filtered,and then fitted to circular arcs by RANSAC.With this method,the calibration is automatically accomplished and image distortion is corrected.In view of the applications scenario of this thesis,the conventional moving object detection and tracking methods are studied and analyzed for selection.Due to the fixed camera position,the block-based Gaussian background modeling is introduced.Then pixel-wise detecting result is calculated statistically.According to the serious distortion of fish-eye camera,an adapted particle filter tracking algorithm based on spherical model is put forward.The experiment result shows that the tracking accuracy can reach 93%,which surpass conventional method.For target tracking in different fisheye cameras,an effective FOV line is proposed and the handover request and response algorithm is designed.The candidates are selected based on the established corresponding relations between two cameras.The intensity difference between two cameras is eliminated by the proposed CCT method and the target describing and matching are based on color features.Similarity for the images of same target in different views reaches 80%,proving the effectiveness of this method.The omnidirectional visual target tracking system is designed using fisheye cameras to expand the monitoring scope of a single node,which eliminates the blind area in traditional methods.The designed system achieves target detection and tracking in single camera,and consistent labeling in the overlapping area of adjacent cameras.Tracking results confirm that the system can effectively accomplish the target delivery task,which comply with the design requirement.