Research And Implementation Of Online-learning Real-time Long-term Tracking

Object tracking is an interdisciplinary field including image processing,machine learning,pattern recognition,statistics and other fields,which has a wide application prospect in video surveillance,production safety,social security,traffic control,military guidance,etc.The correlation filters have recently shown strong robustness and real-time performance in tracking.Using the theory of circulant matrices and the Fast Fourier Transform,KCF avoids the matrix inversion operation in closed-form solutions of the least square method and achieves extremely fast learning and detection;YOLO is an endto-end convolutional neural network which is much more accurate and real-time than the traditional detectors.Therefore,the combination of these two methods is promising in intelligence video surveillance sysems.This thesis proposes two improved KCF trackers: 1.a multi-feature integration and scale-adaptive tracker.2.a long-term tracker combined with YOLO.Both are designed to improve the success rate in various scenes.The main works are as follows:1.FHOG,CN and intensity features were integrated to improve the tracking precision in a variety of scenes.2.Accurate scale estimation was proposed: fast feature pyramid was applied to improve the speed of multi-scale features extracting;the multi-task structure was applied to improve the success rate in scale-variation scenes.3.Learning strategy was improved by updating the denominator and numerator of parameter separately with the previous appearances to avoid overfitting.4.A new CF tracker was studied and implemented with the YOLO feature map.5.A new method of estimating scale which is based on YOLO region proposals was studied and implemented.6.Redetection based on the YOLO proposals instead of sliding window was studied and implemented.The robustness and success rate are evaluated on the OTB database.Among the test trackers,the success rate of hand-craft feature FSKCF tracker in temporal robustness estimate is 62.9%,ranking 1st.The success rate of the CNN feature YOLOCF tracker is 59.2%,ranking 5th.In the scal-variant scenes our trackers are ranked 1st,3rd and run in 80.4fps,30.4fps respectively during 50 sequences.An intrusion detection system was implemented that can automatically track the moving target in complex environment,and achieve the purpose of early warning.