Study On The Object Detection And Tracking Based On Local Feature Extraction

Target detection and tracking as important part of computer vision.It has extensive applications such as motion recognition,automatic monitoring,video retrieval,human-computer interaction and traffic regulation.The traditional target detection and tracking method usually only uses the base features of image,or simple background modeling as the object appearance model.The destination of target detection and tracking is achieved the target location information acquisition.But in the real world,there is object pose and scale variation,illumination,occlusion,complex background and other external environment interference,which is likely to cause useless base features of target image and failed tracking result.Therefore,it is still a great challenge to study a robust detection and tracking algorithm to achieve stable tracking of the target.Extract local features of image to be an important part in the process of object detection and tracking,which directly influences the performance of later detection and tracking.In the aspect of target detection,this paper firstly analyzes the advantages and disadvantages of HOG-PCA descriptor and SVM target detection method.In this paper,we propose a method to improve the HOG-PCA feature descriptor,and use the polar coordinate combination block method to replace the traditional four-quadrant method when extracting the HOG feature.We propose a method to improve the HOG-PCA feature descriptor that can more accurately represent the target.On the other hand,instead of the traditional random sampling method,the positive and negative samples for SVM training are constructed by circulating sampling,which makes the training result of trainer more accurate.In this paper,the improved HOG-PCA descriptor is extracted from the positive and negative samples.And throw the principal component analysis process,after it is used as the sample data to training the support vector machine classifier.Finally,the trained classifier is used for the detection of the target.In the aspect of target tracking target,we also extract the local feature of image and construct the robust target descriptor.In object tracking,the local context surrounding of the object could provide much effective information in getting a robust tracker.The traditional spatio-temporal context learning algorithm is a simple and fast tracking algorithm.However it only use the Gaussian-weighted intensity of image as the object appearance model.The appearance model not enough to deal with complicated tracking scenarios.The paper adopts perceptive features of Circular-Multi-Block Local Binary Pattern(MBs-LBPP,R)feature to construct visual saliency map.And use Spatio-temporal relationship between the target and its local context area modeling on Bayesian framework.In this paper,we propose an object tracking algorithm that use the contextual visual saliency around the object.In this paper,the INRIA pedestrian data set is used to test the SVM target detection method of the improved HOG-PCA feature.It is compared with the original algorithm.The experimental results show that the improved algorithm can reduce the target false positive rate.In the aspect of target tracking target,extensive experimental results on public benchmark databases show that our algorithm outperforms the original STC algorithm and other state-of-the-art algorithms.Finally,this paper summarizes and analyzes the study contents and innovative points comprehensively,figuring out the new thoughts of the algorithm in further investigation.