Research On Tracking Algorithm Based On Semi-supervised Online Learning With Fusion Of Depth Perspective

After years of research and development in the field of target tracking, there have been a lot ofsophisticated algorithms for target tracking. With the development of machine learning theory and practice,the use of tracking algorithm with self-learning mechanism has been steady development. Such detectionalgorithm based within online learning is usually based on an online update of the classifier, generallyregarding foreground objects and background as binary classification problem. It is commonly referred toas target tracking based detection. However, due to the loss of depth information, tracking algorithm basedon color image does not resolve the fundamental problem of occlusion. Meanwhile, the traditional depthcamera equipment is expensive and complicated, it is difficult to meet real-time requirements of targettracking. In2010, the launch of Microsoft’s Kinect depth camera that can simultaneously collect scenedepth information and color information, provides new opportunities for tracking research fields.With target model over-updating in most online learning object tracking algorithm, it often results indeclining tracking performance, eventually tracking failure. To address it, target tracking algorithms basedon online semi-supervised learning framework is the main object of study, such as TLD, Semi-BoostTracker and MILTracker. Based on in-depth study of these algorithms, we propose two object trackingalgorithm based on semi-supervised learning:1. Based on MIL-BOOST framework, we propose a random local mean HASH classifier, to achievethe goal of stable tracking and partial anti-blocking properties of the algorithm is obvious.2. Given the study of random ferns, this paper integrates random ferns and some realtimesemi-supervised learning framework, and proposes a descriptor (HOG-LBP). It implements stable and fasttarget tracking, especially for human tracking.In addition, this paper describes the acquisition methods of depth image, as well as somepre-processing of depth image, such as camera calibration and the depth map filtering. Meanwhile, thepaper describes the nature of depth image and two characterization methods(HOD and Haar-like), proposesa descriptor with color images and depth image(Combo-HOD-LBP), and presents a special target trackingalgorithm within a combination of color and depth images based on semi-supervised online learningframework. The algorithm is based on the thinking of TLD, Semi-Boost and online multiple instanceslearning with the joint processing of depth information and color information. It uses the collaborationwithin each other three modules to achieve superior tracking performance. Meanwhile, an active occlusiondetection procedure based on the target depth histogram changing is presented to ensure that with theoccurrence of the target block, the tracking algorithm can still maintain stable tracking, and avoid updatingof detection classier due to the calibration error.