Research On Transcale Based Moving Object Detection And Tracking

Moving object detection and tracking methods have always been hot issues in computer vision and the key technologies of visual navigation and video surveillance. The main problems of the traditional methods of moving object detection and tracking are the single research scale, the high algorithm complexity and poor generalization ability. They not only ignore the important role of different scales in the movement environment of object detection and tracking, but also result in low accuracy of detection and tracking. In view of the above problems, it has become urgent to research on the problems at present that how to use the transcale characteristics of image sequences for moving object detection and tracking, and how to improve the efficiency and accuracy of the algorithms. In this dissertation, the traditional moving object detection and tracking methods and the transcale technologies are thoroughly studied, and moving object detection and tracking algorithms based on transcale are proposed. The main contributions and innovations of this dissertation are as follows:(1) In view of the problems of the traditional moving object description methods, which only targeted on a single scale space structure,and could not make full use of different scale characteristics of the object,this dissertation proposed a new algorithm to build the transcale space,constructed on Gaussian pyramid and wavelet transform(GWSP). Firstly,the transcale space was proposed through Gaussian pyramid of object due to variation of scale (GWTS) . The high frequency and low frequency characteristics were extracted using wavelet decomposition on every scale. The new transcale space not only considered the details in the image sequences, but also took the object contours into account. Secondly,using the proposed transcale space, dictionary with multi-scale features was trained, and the testing dataset was classified. Experimental results demonstrated the accuracy and effectiveness of the proposed GWSP algorithm under different experimental conditions. Compared with the single scale algorithms, the proposed GWSP algorithm performed better with the trained dictionary on the proposed transcale space. With the trained dictionary, the classification accuracies of the proposed GWSP algorithm respectively increased by 12.8%, 7.1%, 3.9%, 9.6% compared with ScSPM algorithm, MWLP algorithm, HSPMP algorithm and WDSC algorithm, respectively, and the average of them was 8.5%.(2 ) Aiming at solving the problems that most of the current object detection methods were limited to the static images, and the single scale characteristics could not describe the scale invariance of moving object,this dissertation proposed a new transcale based moving object detection algorithm based on three-phase dictionary learning and hierarchial sparse coding (MDSH) . The multi-stage dictionary learning algorithm was proposed, including the initial stage, inter-layer and inter-frame phase, to study and update the dictionary. This learned dictionary was able to utilize the continuously changing new features in video sequences and was more accurate.The hierarchical sparse coding was proposed and used different layers of sparse features to describe the scales of the moving object. Experiments were designed to verify the proposed algorithm for the all-round evaluation, through multiple sets of classification experiments and different types of video sequences. Experimental results demonstrated that, for classification of testdataset, the accuracy of the proposed MDSH method on test dataset increased a lot compared with the existing algorithms. For moving object detection in video sequences,the proposed MDSH algorithm could effectively detect moving object under the environmental interference, achieving better detection results,and the overlap rate was improved by 4.94%. Compared with ScSPM algorithm, on the different numbers of training dataset, the classification accuracies was improved by 1.8%, 10%, 14%, 15%, and by the average of 10.2%.(3 ) Aiming at solving the problem that sampling in the initial stage of the traditional sparse representation method could result in a large number of redundant samples and the problem of shortage of key information, this dissertation proposed a new visual tracking algorithm based on the direction vector and maximum average pooling (DPF-WT).Using the scale features extracted from the image sequences, direction vector was obtained, and applied with sparse representation to combine the global information with the local sparse characteristics, obtaining more robust object features. The proposed feature not only considered the appearance information of moving object, but also combined the displacement and direction in the process of movement. To improve the accuracy of the feature, maximum average pooling method was proposed to obtain the key features of sparse matrix. In order to reduce the influence that the external environments had on the motion estimation,weight selection strategy was proposed, using the reconstruction error of samples to rectify the motion estimation for the observations.Experimental results demonstrated that, through the analysis of the tracking results and evaluation index of every method on different video sequences, the proposed DPF-WT algorithm could be more accurate to track moving object in a variety of environments compared with the existing popular tracking methods, with the center error or the average center error reducing obviously, and the overlap rate and the average overlap rate improving considerably. The average center errors on all the video sequences were reduced by an average of 24.56 pixels, and the average overlap rate was improved by an average of 36.5% compared to that of the existing popular tracking algorithms.(4) In view of the problem of low accuracy and high complexity caused by complex feature extraction methods of object tracking algorithm, this dissertation proposed a naive visual tracking algorithm based on deep learning network and average hashing (DNHT) .Using stack denoising autoencoder, an eight-layer network framework based on neural network was constructed to track the moving object. In training phase, stack denoising autoencoder was utilized to train the weights for each layer of the neural network, and the network was obtained after further fine-tuning. In the tracking phase, low frequency information of every sample were extracted using average hashing method and the scale weights was built to describe the differences between samples and to replace the offset of the network. Thus, through the different characteristics of samples, the feature information of samples with different scales could be extracted. Experimental results demonstrated that the proposed DNHT algorithm could effectively avoid drift caused by challenges of video sequences, such as illumination changes,occlusion, scale, and track the object accurately. The performance in center error and overlap rate was better than the compared existing popular methods, and the average center error on all video sequence was reduced by an average of 15.83 pixels, and the average overlap rate was improved by an average of 34%.(5) Combining the proposed GWSP, MDSH. DPF-WT and DNHT algorithm, this dissertation designed and implemented the moving object detection and tracking system based on transcales (ODTVS) . The ODTVS system included the video frames pretreatment module, moving object transcale describing module, the moving object transcale detection module and the moving object transcale tracking module. With different objective evaluation and subjective evaluation criterion, the algorithm proposed in this dissertation had carried on the all-round analysis and evaluation. In view of the moving object detection, the evaluation criterion were the peak signal-to-noise ratio (PSNR) , the structure similarity index measurement (SSIM) multi-scale structure similarity index measurement (MSSSIM) , iteration error evaluation of algorithm and classification accuracy. For moving object tracking, the evaluation criterion were mainly the center error, the average center error, overlap rate, the average overlap rate, and the success rate.