Object Detection And Tracking Under Low Resolution And Applications

Object detection and tracking is one of the important fields of computer vision. Visual target detection and tracking plays an important role in intelligent transportation systems. This paper focus on low resolution objects detection and tracking algorithm in traffic scene. The main work of this paper is as follows:Firstly, an object detection method is proposed which is adaptive to low resolution condition. In this method, non-negative matrix factorization(NMF) is adopted to build object model. One object is represented by a dictionary matrix which consists of important contours of this kind of objects, and the contours is gotten by iterative calculation based on NMF. Each column of the dictionary matrix is a code word which is stand for a contour. The object is represented as a nonnegative linear combination of these code words. The features of a target are the combination coefficients which are gotten based on the generalized inverse matrix of dictionary matrix. This method breaks the space limitations of the Haar-like feature to describe the target. A row of generalized inverse matrix of the target dictionary represents a filter which gets the weighted sum of pixels to form one dimension of feature. Objects and background parts are spanning by the dictionary matrix, and the features are input to support vector machine classifier for training. In the case of low resolution, the proposed object representation method is robust. When detecting an object, a classic framework of Haar-like feature with Adaboost classifier is adopted to reject most of the parts which are not targets, the remaining candidate regions spanned by object dictionary matrix to get a feature, and input this feature to the SVM classifier for final judgment. The candidate regions which get a positive label by SVM classifier are the detected target.Secondly, this paper proposes the Dynamic Compressive Tracking(DCT) based on Compressive Tracking(CT). In CT, there is a problem that the Random Measurement Matrix(RMM) is static and can be updated. DCT uses a Dynamic Random Measurement Matrix(DRMM), and could be updated dynamically according to the movement of the target to discriminate the difference between the target and background. This paper defines the framework of the DRMM, proposes two necessary conditions to choose the classifier to update the DRMM. CT and DCT are tested on two public dataset. In tracking, the feature that can’t discriminate the target and background is replaced by a random new one, because of the update of the RMM. This method reduces the influence of the bad feature, and proves that DCT is better than CT.A vehicle detection and tracking subsystem of the electronic police system is built to verify the low resolution object detection and tracking algorithm. In detection part, the framework of Haar-like feature and Adaboost classifier is utilized to reject most of the non-target region, and the rest region is input into the SVM classifier for verification. DCT is utilized in tracking part. When search the new location of the target, the scale is considered. This system solves the problem that the traditional electronic police system can’t detect and track the low resolution target far away.