Improved Differential Evolution Algorithm And Its Application In Object Tracking

Computer Vision,which is a research field of great concern,has been developed rapidly recent years.It is a fundamental research subject of Computer Vision to detect and track specific targets in the image.An effective tracking scheme can enhance the machine intelligently “perceive” the image and make a better “decision”.The process of tracking can be viewed as a dynamic optimization,namely the location of the best area of the target in image sequence.As a new kind of optimization algorithms in intelligent computing area,differential evolution(DE)has been widely studied and applied.DE algorithm is simple,parallel and self-adaptive when dealing with several problems in science and engineer.Because of some shortages in DE such as stagnant evolution and premature convergence,its optimal ability is limited for specific applications.For better adaptability in object tracking algorithm,we make some improvements in DE referring to current literatures.Firstly,we introduced two discarded individuals during the mutation stage which further enrich the individuals' diversity and speed up the evolution.Then,in middle-late stages of evolution,for mining more valuable information of these discarded individuals,these individuals are accumulated as aided population and its population size can be adjusted along with the original population.Experimental results also show that the improved algorithm has fast convergence speed and higher accuracy.In specific tracking application,in order to deal with the complex tracking scenario,we carry some image preprocessing and build background model based on adaptive Gaussian Mixture Model(GMM)for better real time and robustness.In the end,experiments demonstrate that the designed tracking algorithm is successful in most video datasets and many complex tracking scene and interference factors can be handled effectively.