Research Of Object Tracking Techniques Based On Machine Learning

As an important research topic in the computer vision,object tracking has been widely applied to intelligent monitoring,industrial automation,intelligent transportation and other fields.In recent years,machine learning techniques have been applied to object tracking.They regard the object tracking problems as binary classification problem of the object and the background.But in the complex environments,the object tracking based on machine learning still has a lot of difficulties and challenges,such as the changes of object and background,illumination variation,object occlusion,the fast moving and deformation of the object.To solve these problems,the main contents and contributions of this paper are as follows:1.Representation of the object feature.The traditional image features have the problems of complex calculation,incomplete description of the object,lack of stabilities in illumination and deformation.This paper uses histograms of oriented gradients(HOG)for image representation.Compared to other features,HOG can describe the local information of the image well and has translation and rotation invariance.In addition,HOG feature is robust to illumination variation and local geometric distortion.2.The design of classification model.How to design an efficient classifier to enhance the performance of the tracking algorithm in complex background and motion blur scene is still a challenging problem.In this paper,a single hidden layer neural network is used as a classifier and extreme learning machine(FLM)algorithm is used to train the classifier.Compared with the traditional classifiers,the ELM algorithm has the advantages of good generalization ability,not easy to fall into local extremum.3.Object search mechanism.In order to solve the problems such as fast motion of the target,this paper proposes an object search mechanism that based on the multi-scale sliding window.Firstly,the motion model of the object is established and the object candidate region of the current frame is predicted by the model.Then the algorithm carries out multi-scale sliding window detection within the candidate area.4.Research of the classifier online updating.The traditional object tracking algorithms cannot adapt to the changes of object and background in the large space.Based on this,this paper proposes an online updating algorithm based on ELM.In the tracking process,regions around the object are selected as positive samples and regions in background are selected randomly as negative samples by using the tracking results in the current frame.Then the positive and negative samples are sent to the classifier for learning(updating)the parameters of the classifier.The object tracking algorithm proposed in this paper is validated in the Visual Tracker Benchmark database.The database collected a large number of object tracking videos which contain several of the above tracking difficulties.According to the experiment results,the object tracking algorithm based on ELM that proposed in this paper can effectively solve the problems existed in the object tracking process,such as occlusion,complex background,illumination variation,motion blur,deformation and fast moving,etc.