Research On Object Tracking Algorithms Of Classifiers Based On Multi-instance Learning

Target tracking is a popular research direction of computer vision today,and it is widely used in UAV surveying,autonomous driving,security monitoring,traffic control,national security military and other fields.Although target tracking has been successfully applied in many systems,it still faces many challenges,such as changes in illumination,occlusion,deformation,motion blur,complex background,and low video resolution during the tracking process,which will cause tracking failure.Therefore,a design is designed.A target tracking algorithm with strong robustness and high accuracy is the research focus.The classifier in the target tracking system is mainly used to distinguish the background and the target,so as to determine the current state of the target.Here,the multi-instance learning(MIL)-based classifier in the target tracking system is improved to improve the robustness of the tracking algorithm The main research contents are as follows:(1)Aiming at the drift problem that may occur when the multi-instance learning target tracking algorithm updates the classifier,which leads to poor tracking effect,a MIL target tracking based on information entropy feature selection and example weighting is proposed.First,use the multi-instance learning method to collect positive and negative samples around the marked target position in the video initial frame,that is,the background image block and the image block around the target,and assign different weights to the positive and negative samples.The method of calculating the sample weight adopts SSIM to extract The obtained image features form a weak classifier set through online learning.The extracted image features are compressed in combination with the compressed sensing theory,and then several optimal weak classifiers are selected from multiple weak classifiers according to the principle of maximum entropy to form a strong classifier.Input the next frame of unmarked video image,continue to collect background samples and target samples,and perform the second classification of the background and target on the samples collected in the new frame through the strong classifier generated above,and the final image block with the largest probability value is the target Image block,that is,the target location.Experimental results show that the algorithm exhibits good robustness under the interference factors of target occlusion and motion blur.(2)Aiming at the re-detection module of the kernel-related filtering target tracking algorithm,in order to further improve the robustness of the algorithm,a target tracking based on mi-SVM and kernel-related filtering is proposed.mi-SVM is an SVM classifier that introduces the constraints of MIL.Initialize the first frame of the video input,and the subsequent video frames will crop the image block at the previously estimated position as input,and then extract the image features.After training,the spatial confidence map is obtained,that is,the feature calculated by the kernel correlation filter.The response value,the maximum response value is compared with the activation threshold ??re in the re-detection module.If and only if the maximum response value is less than ??re,the re-detection module is activated,and then mi-SVM is used to recalculate the target position.The calculated result is if If the response value is greater than the original value,the result of the re-detection is used as the new position of the target,and finally the current position information of the target is used to train and update the kernel correlation filter.Experimental results show that the algorithm performs better in terms of fast motion and motion blur.This paper uses the classic OTB100 data set for testing,and selects seven classic target tracking algorithms for comparison.The experimental results show that the MIL target tracking based on information entropy feature selection and example weighting improves the overall performance of the original MIL by about 50%.The target tracking of mi-SVM and kernel correlation filtering is about 15% higher than that of the original KCF.Therefore,multi-instance learning has a relative advantage in the target tracking system.