Research On The Improvement Of Self-organizing Background Subtraction And Object Tracking Algorithm

The self-organizing background subtraction(SOBS)algorithm and the tracking algorithm based on color distribution are the classic object detection and tracking algorithms.Aiming at the drawbacks of the original algorithm,this paper proposed some new ideas,and designs an improved algorithm to improve the precision of detection and tracking results.First,considering SOBS algorithm suffers from some drawbacks:poor real time performance and the background model being prone to bias under complicated environment due to non-periodic changes,an improved algorithm is proposed.Firstly,according to our experiment,the background of SOBS algorithm model while using a large number of nodes to record the status of the pixel,but the redundancy between nodes is very high,so this paper constructs a new boundary sharing model to reduce the space complexity.Secondly,to deal with the "shift" of model,this paper designed and introduced a structure called "memory organization",used to store the special nodes in history.When the“shift”occurs,the memory will be called to correct the model.Finally,SOBS algorithm requires the background in a weak change,which limits its application environment.In this paper,by introducing the background transfer decision mechanism,the detection algorithm can cope with the global change of the background.At the same time,the use of memory storage organization can make the background model quickly recall the previous scene.Second,aiming at the limits of the traditional color distribution model,an improved extended color distribution model is proposed.First of all,we design a kind of "nested distribution" expanded model structure,in order to enhance the ability of color expression.Secondly,the traditional color distribution model in the calculation of color distribution,assuming that the importance of color is related to its location,and many practical problems can not meet the hypothesis,resulting in the tracking effect greatly reduced.In order to overcome this shortcoming,this paper designs a new method to calculate the color distribution of the object according to the background's color distribution.A large number of experimental results show that the new extended color distribution model can effectively improve the ability to resist occlusion and background changes.Third,a new tracking algorithm based on the color model and HOG feature is proposed.In the feature description,our algorithm combining color and texture information:firstly,using the color model just mentioned above as the object's color feature;second,using HOG features to make up for the inadequacy of the color feature.The algorithm uses particle swarm optimization algorithm to get the object's location,this kind of random search method with the guidance can reduce time complexity as well as avoid the search into the "optimal" trap.A large number of comparative test results show that the tracking algorithm proposed in this paper is better than the popular tracking algorithm in most cases.Fourth,aiming at limits of the traditional detection and tracking system(search range is limited and the scene has to be fixed),this paper proposed a random object detection algorithm based on the division of vision to extract the object which is attractive,and on this basis a tracking system framework based on the perspective of semi fixed is put forward for practical application.