Research On Stable Detection Technology Of Extended Objects

As a significant component of image processing,object recognition and tracking,object detection has always been a hot research issue.It has a wide range of applications in many fields such as video surveillance,vehicle navigation,object tracking and so on.Traditional methods for object detection are mainly based on feature point detection,however,there is some problems in the complex scene such as high leakage rate,low detection rate and so on.This paper is mainly focused on extended object detection,usually such objects occupy a large position in the view filed,and simultaneously,they have many other problems such as larger posture,bigger scale variety,easily obscured and so on.However,traditional detection algorithm based on feature point matching probably leads to the feature matching failure and affects the detection rate.In addition it also has an influence on follow-up applications such as object tracking.Therefore,proposing a more robust and stable object detection algorithm is very significant.For extended target detection,this paper makes a detailed analysis of the local feature extraction methods commonly used include SIFT,HOG,Harr-like,LBP,and HOG features are analyzed and improved in depth.Because of the problem that HOG can't adapt to the change of the object scale,construction of the feature pyramid make the HOG features have the property of scale invariance.In order to solve the problem that the feature dimension of HOG is too large and the computation is large,principal component analysis(PCA)is used to reduce the dimension of feature.Using support vector machine(SVM)to train the sample image,to achieve the HOG+SVM target detection algorithm.Aiming at the problem that the target is occluded and the posture is changed,it is easy to cause the missing and false detection,and the target detection algorithm based on the part is constructed from the angle of the part.In the process of training the model parameters of the part,because the input sample only labeled the whole object and did not label each part,Latent-SVM training is used.In order to further improve the robustness of the algorithm,in the training process of the model parameters,firstly the samples were clustered,according to the clustering results,a multi perspective mixed training model is built,and realize the target detection algorithm of multi components.The experimental results show that the algorithm can detect the target accurately when the object is deformed,occluded and changed.However,although using the component concept to detect object can improve the detection accuracy,accompanying the increase of model complexity,the detection speed slow down.To solve this problem,the idea of cascade detection is employed in this paper,by the cascade clipping algorithm,the image areas without the object can be quickly filtered out.The experimental results show that with the conception of cascade detection,the method can improvethe computation efficiency and shorten the detection time while preserving the same detection accuracy.