Research On Object Tracking Based On Semi-Supervised Learning

Object tracking is one of important research topics in the area of computer vision.It depends on various research methods, such as pattern recognition, machine learning, image processing etc. Object tracking algorithm is divided into the object tracking algorithm and the object detection algorithm, based on whether or not there is a prediction mechanism included in the algorithm. Detection is utilized to extract the foreground in the image. With the great development of the computing power and storage capacity, the detection algorithm has become a important research direction in the area of object tracking.This work focuses on the implementation of detection algorithm in the area of object tracking. In the presence of background clutters, occlusions and changing illuminations etc during object tracking, a cascade classifier is designed aimed to handle such a complicated situation. Lucas-Kanade optical flow is utilized to predict an approximate area so that the running time of algorithm is reduced. The major contributions of our work are as follows:Firstly, a cascade classifier is designed, composed by a variance filter, a random forest classifier, template matching and cluster analysis. An inputted frame is processed into patches by sliding-window, and then patches are inputted into the cascade classifier one by one. Only if patches meet the conditions of the current weak classifier, they will be sent to the next step. The functions of each model are as follows:(1) The variance of an image patch is a measure method for the uniformity of patch and is able to eliminate background patches rapidly; (2) The random forest classifier uses the simple 2bitBP to detect the patches, in order to increase the speed of algorithm; (3) The template matching checks the patch in every pixel with templates and outputs a final label of the patch; (4) The cluster analysis fuses all the patches that go through Template matching and outputs the tracking object’s final position.Secondly, the self-learning model is introduced in our proposed scheme, in order to handle the challenges caused by various complicated situations, such as the occlusions or objects’ leaving out of the camera view. The functions of self-learning module are as follows:(1) The outputs are checked again to correct the wrong judgments about the labeling of patches. (2) The corrected patches are added into the training set to train the cascade again and also added to template set. In addition, we design a template update strategy. A threshold of template number is set to limit template update to reduce time spent on template matching.Finally, LK optical flow algorithm is used to estimate an approximate area so that the overhead brought by the global searching is reduced. LK optical flow uses the result of the last frame to predict an possible area of the object in the current frame so that the running time of our algorithm is reduced.Experiments are done on several sequences of test videos to verify the efficiency of our algorithm. It demonstrates that our algorithm can deal with the challenges caused by the complicated situations, including background clutters, occlusions or illuminates, and its performance is better than some existing classical object tracking algorithms.