Research On Object Tracking Technology Under Complex Background

This paper makes a research on object tracking under complex scenarios, aiming at solvingthe problems of computing speed and tracking stability when the object changes its appearance,gets occluded etc.. Tracking methods based on online classifier have advantages on handling theobjects’ appearance changing, occlusion and re-capturing the object after it gets disappeared. Soour research is mainly based on this kind of methods. And we make some improvement on theshortcomings.As a classic tracking algorithm based on online classifier, online boosting algorithm onlyuse the detection result of every frame as a positive sample to update the classifier. If there weresome errors in the detection, the classifier’s performance would get decreased in the next frame.To solve this problem, in the multi-instance learning algorithm, multiple instances are sampledaround the detection position and combined to form the positive bag. Even there existed someerrors in the detection result, the real object is ensured to be included in the positive bag, thusreduce the cumulating of the labeling error. The large number of positive instances makes thecomputing of the bag log-likelihood time costly. So we approximate the sample bagslog-likelihood function and use the gradient descent method in the function space to simplify thetraining process, thus dramatically improve the classifier’s training speed. Traditional Noisy-ORprobability model can’t characterize the negative sample bag well when its capacity is large, sowe use the geometric mean model instead. The experiments show that our model has a betterstability.Rectangular bounding box is commonly used to represent samples in the online classifierbased tracking methods, which inevitably introduce some background component to the positivesamples. In order to avoid the interference of background in the positive sample, we presents amethod combining level set segmentation to divide the object area to ensure only the charactersin the foreground in the positive samples are learned by the classifier. Experimental results showthat our method can inhibit the interference of background components in some degree. Slidingrectangular window brings another defect to online classifier based tracking methods: have pooradaptability to scale and rotation changes. To solve this problem, we proposes a Bayesianframework which combine the target location, scale, rotation angle and shape informationtogether, the tracking problem is treated as the joint estimation of the above information, and wesolve it by maximizing the posterior probability. This method completes the segmentation and tracking tasks at the sa me time, get pixel-wise precision and adapt the scale and angle changewell.