Design And Implementation Of Visual Tracking System Based On Kernelized Correlation Filter

The autonomous tracking of mobile robots is an important research direction,and the following technologies have broad application prospects in life and production.In this thesis,a high accuracy,high real-time and high robust target tracking method is designed.A visual tracking system is built on the mobile robot GH_ROBOT experimental platform to complete the target pedestrian follow-up.In this thesis,an improved method of kernelized correlation filter tracking algorithm is proposed.In the background of complex interference,the target pedestrian is perceived and the foreground tracking object is automatically initialized.The improved kernelized correlation filter is used to track the region where the moving target is located to realize a low cost,accurate and real-time autonomous following system.The main work includes: 1.Improve the kernelized correlation filter tracking algorithm.The FHOG features and color features of the sample are linearly blended to enrich the target appearance features;the scale pyramid is introduced into the kernelized correlation filter to solve the problem that the target changes in the course of the motion leading to the following failure;The APCE is used to judge the reliability of the current frame tracking to perform model update of the position correlation filter.2.Design based on improved kernelized correlation filter visual tracking module.The SVM classifier is trained according to the pedestrian sample set,and the pedestrian is automatically detected as the initial frame of the tracking algorithm,and the target re-detection is introduced.3.Design a visual tracking system for mobile robots.According to the motion model of the mobile robot,the relationship between the motion speed and the chassis driving wheel is obtained.The motion control based on the speed difference makes the motor speed and the command speed approximately the same;the visual servo control makes the target pedestrian in the central axis of the mobile robot and stay within a certain security threshold.The visual tracking system is implemented on the built mobile robot GH_ROBOT experimental platform.The visual tracking system includes three nodes: RGB-D sensor,target detection and target tracking and lower computer controller.The test results using the OTB-100 data set show that the improved kernelized correlation filter algorithm proposed in this thesis has good accuracy,robustness and real-time performance in dealing with complex tracking scenarios.Compared with the typical kernelized correlation filter,the optimized algorithm has improved the average tracking accuracy by nearly 15% and the average tracking success rate by nearly 33%.Experiments on the visual follow-up system of mobile robots are carried out for the morphological changes,illumination changes,target occlusion,scale changes,and out-of-field interference.The reliability of the improved kernelized correlation filter tracker and re-detection mechanism for target pedestrians for a long time is verified,and the accuracy of follow-up exceeds 90%.