Human Head Tracking Based on Inheritance and Evolution Concept

This paper presents a method for tracking human head in cluttered scenes to achieve robustness to occlusions and environment change by introducing inheritance and evolution concept into tracking system. Different from most works on tracking where detection and tracking are loosely coupled, we view the essence of object detection and tracking as a process of modeling object class classifier (in detection stage) and object instance classifier (in tracking stage), they are coordinated in the class/instance relationship. At first we trained a head generic class classifier in off-line learning, which is a hierarchical structure where selected features are arranged from rough to detail and from low to high level. Then we derived the tracker’s instance classifier from the base generic head classifier. The instance classifier is also a hierarchical structure. Its features in low and middle levels are initialized by inheriting low and rough features from generic base class classifier. Its high level features, the individual identification parts, are learned and evolved on-line with boosting method. By incorporating inheritance and evolution into tracking, the tracker not only adapts itself to the surrounding change, but also gains the ability naturally to distinguish it from other instances of the same class. The experiment shows its effectiveness.