Research Of Object Tracking Based On Random Forest

Video target tracking technology is an important research direction in the field of computer vision, which is widely used in important areas such as intelligent transportation, smart home, security monitoring. Random Forest as a method in the field of pattern recognition has been rapidly developed in recent years, due to its speed, high-accuracy, parallelism, etc. And it has been widely used in the field of character recognition, image classification, but for tracking the method is also unusual.This thesis first carefully describes the theoretical basis of the Random Forest, from the formation of single decision tree to building Random Forest. Then we introduced two random forest improvements, one is transforming offline Random Forest into online version, making it suitable for online video tracking; the other is that introduce generalized Hough transform theory to Random Forest method, making it more suitable for fragment video tracking.Then we put forward two tracking framework, the holistic framework based on online Random Forest and local framework based on online Hough Forest. In holistic framework,we first put forward an online weighted Random Forest algorithm which combines with Adaboost;then we take overall target as the tracking target and introduce co-training theory to improve tracking accuracy. And in local framework, we break the overall target into fragments, and take full account of the position information of the pieces. We determine the overall target position through the position of pieces. Experimental results show that these two methods can achieve accurate tracking in some video, but there are some defects also.Finally, we analyze the advantages and disadvantages of the holistic and local tracking separately, and we put forward a new framework which combines the holistic tracking and local tracking. The experiments show that this method can overcome some difficulties which in the separate tracking methods.