The Study Of Object Detection And Tracking For Transcale Motion Images

Transcale spatial object detection and tracking make use of the spatial motion images, captured by sensors, to analyze spacecraft flight orbit and space station capabilities. The final detection and tracking results could provide valuable reference informations for spatial defense, aerospace motion and space station operations. Limited by spatial environmental factors and the defects of spatial camera devices, spatial motion images inevitably appear weaken effects in the process of capture, transmission and storage. These weaken effects cause large detection errors, reduce the authenticity of spatial object tracking, limit spatial docking and interfere spacecraft surveillance.This dissertation studies the transcale spatial image feature extraction, transcale spatial object detection, transcale spatial object tracking. The contributions and innovations are as follows:(1) This dissertation proposes a Bayesian Non-Local Means filter (BNL-Means), this filter could extract spatial motion image feature effectively. This dissertation proposes a Spatial Image Transcale Feature Extraction(SITFE) algorithm, which uses Non-Subsampled Contourlet Transform(NSCT) to complete the transcale decomposition. The dual filter composition of NSCT could solve the single-translation problem in2-D filters and its associated mapping method can remove the additional data redundancy. The experimental results prove that, compared to the existing filters, BNL-Means could improve Peak Signal to Noise Ratio(PSNR)1.8%and increase Mean Structural Similarity(MSSIM)2.91%. Under different noise levels, SITFE could improve PSNR1.5dB-1.9dB compared with Curvelet feature extraction algorithm and improve PSNR1.0dB-1.2dB compared with Contourlet feature extraction algorithm, under different types of noise interference, SITFE could increase MSSIM18.9%compared with Curvelet feature extraction algorithm and increase MSSIM12.1%compared with Contourlet feature extraction algorithm.(2) This dissertation proposes a Wavelet Optical Flow(WOF) algorithm. The WOF algorithm could accurately estimate different moving objects in the same scene and solve the traditional optical flow estimation problem that moving objects with high speed would be detected less accurately, the WOF algorithm also reduces the optical flow calculation complexity and improves the efficiency of optical flow estimation. This dissertation proposes a Rectangle Window Scan(RWS) algorithm, the RWS algorithm could realize adaptive object detection and support continuous object detection between interval frames. The experimental results prove that, compared to Lucas Kanade algorithm, Horn Schunck algorithm and Occlusion Aware Optical Flow algorithm, the WOF algorithm could increase SFDA10.4%,13.6%and12.7%respectively and improve the computational efficiency28.94%,27.65%and38.11%respectively. Compared to the existing object detection methods SIFT, BS and HF, the RWS algorithm could improve the Object Detection Accuracy(ODA)7.29%,16.57%and9.35%respectively, increase Object Detection Precision(ODP)3.41%,17.5%and4.97%respectively.(3) This dissertation proposes a Particle Filter Object Tracking(PFOT) algorithm based on Re-Sampling Particle Filter(RS-PF). The PFOT adds adaptive multidimensional signal processing, which could dig out new varieties from the neighboring particle clusters and assure that the final object tracking results would not be affected by sampling position, object speed rates and other interference factors. PFOT algorithm uses3D Fourier transform to deal with signal parameters. In3D colorful Fourier domain, PFOT algorithm combines color harmony with signal processing to achieve initial object segmentation and uses Re-Sampling Particle Filter(RS-PF) to complete object tracking. The experimental results prove that, compared to M-PF, S-PF and R-PF, the PFOT algorithm could improve particle sampling accuracy10.35%,27.36%and29.05%respectively. Compared to the existing object tracking methods SDSROT and ILRVT, PFOT algorithm could reduce the object tracking Sum Squared Errors(SSE) over31.21%and45.78%respectively.(4) This dissertation proposes an Active Contour Object Tracking(ACOT) Calgorithm, ACOT algorithm could achieve the real-time update of moving object characteristic regions and their adjoining region, it could robustly detect the shape characteristics of moving objects which have sudden changes. ACOT algorithm integrates background modeling and object shape boundary information, it could achieve adaptive object tracking based on moving object shape contour. The ACOT algorithm could overcome the topology limitations of the existing contour detection algorithms, identify the active contour of moving objects and improve the tracking accuracy. The experimental results prove that, compared to RTSTS algorithm, ACOT algorithm could reduce the Root Mean Square Error(RMSE)32.32%in the X-axis and reduce the RMSE20.41%in the Y-axis, compared to SAP algorithm, OACM could reduce the RMSE20.96%in the X-axis and reduce the RMSE15.19%in the Y-axis. Compared to RTSTS algorithm and SAP algorithm, the ACOT algorithm could reduce the object tracking Euclidean distance21.67%and21.26%respectively.(5) This dissertation proposes a Monte Carlo Contour Evolutionary(MCCE) algorithm, MCCE algorithm uses prototype pyramid perception transform and Monte Carlo contour detection to achieve the optimal spatial motion image representation, the Monte Carlo contour detection could approach spatial moving object at different transcale layers and extract moving object feature between different scales. MCCE algorithm could exclude spatial background interference to the maximum extend and achieve precise spatial motion image contour detection. This dissertation proposes an Enhanced Singular Mean Shift (ESMS) algorithm, ESMS algorithm uses the strengthen singularity mean shift calculation to achieve visual enhancement in spatial visually sensitive region. ESMS algorithm makes use of affine box parameter estimation to achieve adaptive spatial motion region adjustment and exclude spatial information interference. The experimental results prove that, compared to Canny edge evolutionary algorithm, MCCE algorithm could reduce the contour detection MSE15.46%and increase PSNR15.63%, compared to Laplacian edge evolutionary algorithm, MCCE algorithm could reduce the contour detection MSE8.72%and increase PSNR4.62%. Compared to ODML algorithm and MAMS algorithm, ESMS algorithm could reduce the object tracking Euclidian distance16.9%and17.8%respectively. Compared to MAMS algorithm, ESMS algorithm could reduce the object tracking MSE23.06%, compared to ODML algorithm, ESMS algorithm could reduce the object tracking MSE25.24%. Compared to MAMS algorithm and ODML algorithm, ESMS algorithm could reduce the object tracking Bhattacharyya distance29.7%and22.5%respectively.