Self-constructive multiple object detection and tracking in unstructured dynamic human environment = 비구조화된 동적 인간 환경에서의 자가구성 다수 물체 검출 및 추적

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 405
  • Download : 0
In this thesis, a novel self-constructive point-level multiple object tracking (PMOT) method is proposed for autonomous robots operating in unstructured dynamic environment (UDE). With the particular assumptions in UDE, no prior knowledge of target objects, dynamic change in their shapes and numbers, and using the sensor system only on the robot, the proposed method features a self-constructive object file and the assimilation and accommodation processes, that are based on the Piaget’s cognitive development theory, in order for interpreting the observed point-cloud data, robust tracking, and incremental learning the target object models. The three components are designed and implemented to achieve the three objectives: flexibility in representing priorly unstructured multiple objects, reliability in tracking dynamic multiple objects that are interacting each other, and computational efficiency for real-time implementation and extendability of the proposed method. The self-constructive means that not only the parameters but also the structure of a model can be learnable and reconstructable according to the change of environments. In order to represent arbitrary unknown objects, an object is embodied as a Gaussian Mixture Model (GMM) that is a probability density function showing the object’s existence stochastically in 3-d space. The proposed self-constructive object file is designed as a hierarchical structure, where multiple objects are stored in point-level, component-level, and object-level. The component-level data represents each Gaussian of objects, and allows the self-constructive property by constructing spatiotemporal association among them incrementally. Based on the Piaget’s cognitive development theory, the assimilation and the accommodation processes are designed to achieve reliability in identifying multiple objects and updating each object model for representing each object in the object file. In order to achieve the research objectives, the pro...
Kwon, Dong-Sooresearcher권동수
한국과학기술원 : 기계공학전공,
Issue Date
568422/325007  / 020095007

학위논문(박사) - 한국과학기술원 : 기계공학전공, 2014.2, [ iv, 102 p. ]


Robotics; RGB-D 카메라; 다수 물체 추적; 시각추적; 기계학습; 로봇비젼; Human-Robot Interaction; Robot vision; Machine learning; Visual tracking; Multiple object tracking; RGB-D camera; 로봇공학; 인간로봇상호작용

Appears in Collection
Files in This Item
There are no files associated with this item.


  • mendeley


rss_1.0 rss_2.0 atom_1.0