Research And Implementation Of Moving Target Classification In Intelligent Video Surveillance Platform

Intelligent video surveillance platform is an important research topic in the field of machine vision. With the rapid development of economy, science and technology\ society are growing, the living standards of people also will be improved. Thus the demand for intelligent video surveillance industries are growing, which makes it more and more scholars to research intelligent video surveillance platform. The main role of intelligent video surveillance platform is detecting moving object, classification, tracking and behavior understanding in the surveillance video. Among them, target classification as a critical step in building intelligent video video surveillance platform, its main purpose is the early detection of moving targets out of interest are classified, for the latter to continue to provide information to track and understand behavior.This paper is the study and implementation of moving target classification in intelligent video surveillance platform. Algorithm used herein is the use of probabilistic decision tree classification algorithm framework with Adaboost. The main steps of the algorithm is the first of a variety of features and feature extraction schemes studied, as a comparison to determine the Haar-like feature of the algorithm is required. Then all training sample set of positive and negative will be extracted feature. The extracted feature values are for Adaboost training. Make the every trained strong classifier as a probabilistic decision tree node, and then continue to the subset of Adaboost training, the formation of a child nodes until the error rate is zero or less than the number of samples 20 end of the training, constitute the probability tree by tree each node. Finally, through the decision probability calculation, enabling the moving target classification and recognition in the video scene.This paper combines Adaboost with probabilistic decision tree(PBT) framework, which is different from traditional Adaboost classification algorithm. Each node of the probability decision tree is a trained strong classifier. Classification results continue then classified until you get the correct classification. In addition, the paper also proposed to calculate the posterior distribution after used to complete the sentence for each node classification. Though the complexity of the algorithm increases slightly, the uncertainty of results are reduced, and the accuracy of the algorithm are improved.