Study On Algorithm Of Multiple Objects Tracking In Complicated Scenes

Recent years, the applications of computer vision have been popularized constantly. Thecomputer vision products such as intelligent surveillance system, vehicle assistant drivingsystem and industrial robot have facilitated people’s life and increased productivity. Computervision is comprised of two layers. The lower layer is the background processing, objectdetection and tracking, while the upper layer is responsible for object recognition andclassification, behavior comprehension. Object tracking as the lower layer of computer visionsystem plays quite an important role and relates to the system directly. Currently, in order tosolve the problem of video tracking, researchers propose many solutions from differentperspective. Object tracking algorithm could be divided into two kinds. One is the trackingalgorithm based on the data filter, and the other one is tracking algorithm based onclassification. For the first one, tracking is treated as solving the problem in the state space. Itpredicts the state as the Bayes theory and solves the state equation to fulfill tracking. KalmanFilter and Particle Filter are two typical examples of data filter based tracking. For thetracking algorithm based on the object classification, classification algorithm is the core. Itconsiders tracking as a binary classification problem. Tracking algorithm discriminates theobject from the background. The classifiers are selected from a feature pool by ensemblelearning. The traditional classification algorithm as the off-line learning is hard to trackobjects well. The main reason is that the classifier is trained on the dataset adopted in advance.Once the classifiers are trained, they are fix and could not adjust according to the datagenerated with time. Due to the object is usually uncertain and change by scale andperspective, the result of classification is not ideal. However the on-line learning algorithmovercomes this disadvantage, classifiers could be updated incrementally with the datagenerated with time. It tracks object in the general scene. We summarize and analyze the basictheory of the tracking algorithm based on the data filter and classification in this paper. We introduce the Bayes filter theory, Kalman filter theory, Particle filter theory and OnlineBoosting algorithm. By analyzing and comparing the advantage and disadvantage of differenttracking algorithms, we propose a tracking algorithm combining object motion informationand classifier based on Haar feature. Kalman filter as a filter model is used to model themotion information. It is used to predict the possible position in the next time. OnlineBoosting algorithm can model the Haar feature and update the classifier on-line with the newsamples generated as the object moves. The performance can be guaranteed by online learning.This algorithm predicts object priority position with Kalman filter, then it searches a localregion around the priority position for the area of the highest confidence with classifier basedon Haar feature. The area is used as the measure value of Kalman filter, then a Kalman correctprocess is adopted to obtain the optimal estimate. Finally our algorithm updates classifier withonline boosting algorithm. In the end, we do the experiments and compare the our algorithmwith the tracking algorithm based on Online Boosting. Experiment results indicate that ouralgorithm is capable of tracking object steady and tolerating the change of object shape. Ouralgorithm is superior to traditional tracking algorithm based on Haar feature classifier incondition that tracking object is covered by objects of the same category. Due to the timecomplex is relatively low, it could be the lower layer for computer vision application.