Real-time compressive tracking was a simple and effective tracking algorithm. However, there were a number of problems which need to be addressed. First of all, it was easy to introduce errors due to factors such as occlusion and clutter. Secondly, it couldn’t update the positive and negative samples accurately while using fixed tracking window. At last, the number of testing samples was too large, which affected the speed of tracking. The occlusion was checked by comparison between consecutive frames’ histograms, and the coefficient can be also updated adaptively by the comparison result. We searched for more specified areas with multi-scales to find out the best matching place, and to handle scale change of the target on the basis of the original algorithm’s tracking result. The different numbers of sub features sets were utilized to filter the testing samples. In that case, the speed of tracking process would be improved. The strategies we proposed would improve the original algorithm’s performance to avoid the failure of tracking. The experimental results indicate that the algorithm can run in real-time and perform favorably against state-of-the-art algorithms on challenging sequences in terms of efficiency, accuracy and robustness.

In the traditional photoelectric tracking system, the fixed scale filter is usually used for target detection. The target features especially the target size is changing all the time due to the distance change between the target and tracking system. It results in the filter not suitable for the target size changing, and it is difficult for the system to obtain the best extraction results. For this question, a small target detection method based on adaptive scale is proposed. Based on the Laplacian scale-space theory, the normalized Laplacian scale-space images are studied, the target signal is enhanced and the noise is suppressed at the same time. By finding the best parameter of target both in the scale and space, the small targets with different scale could be detected effectively. The simulation test and the experimental results show that the method proposed improved the detection capability of the system for small target with different size remarkably.

Moving object is very difficult to be associated with the time and space elements, and also difficult to solve the association. To solve the problems, a algorithm is proposed which take optimal path set’s data association algorithm, and the discrete particle swarm algorithm is brought to the solution. Firstly, the algorithm fuses the object’s appearance match similarity, time and space constraint by Bayesian network net, and then transforms the data association problem into the optimal path choice in the network. Secondly, the average sample pairs’ similarity between paths is set as evaluation function, and when its value is the maximum, the path is optimal. Finally, the emergence of target is discrete on time and space elements, so the discrete particle swarm algorithm is used to obtain the optimal path, and the target’s moving path is recorded by particle encode. The algorithm makes a tracking target simulation in the networks composed of five cameras. The results show that it can effectively solve the multiple target optimal path set, get the target’s movement in the network, realize the relay track, and have good robustness.

Aiming at the problem of leak detection, hollows and false targets in the traditional frame difference algorithm, an improved frame difference algorithm is presented. The effect of moving target detection is experimented and analyzed in several common color spaces, selecting excellent color channels to build hybrid color space CbVb* for moving target detection. In order to make full use of the correlation information among frames, seven frame difference algorithm is presented to calculate moving target frame difference in time domain according to the pixel changing characteristics of CbVb*, and edges of moving target in gradient domain is obtained by using Canny operator with adaptive threshold. Then, frame difference in time domain combines with target edges in gradient domain, and the combined information is carried on corrosion and expansion processing to get the final detection results. The experimental results show that the improved algorithm can more accurately detect moving targets, and have good robustness, adaptability and real-time performance.

On moving platform with little space, theodolite is apt to lose laser cooperation object and appear laser diffuse reflection phenomenon leading laser measurement distance information invalid, which is because of the short intercept and fast tracking angular velocity in the target approach segment (from the flying targets <200 m). In order to solve the problem that non-laser ranging can not use the original method to measure trace, a method of precise measurement by single theodolite on moving platform without laser ranging is presented based on the principle of spatial geometry and collinear condition, the combination of theodolite accurately calibration, midline video, theodolite video and motion attitudes of moving platform. Test result shows that the method has the advantage of simple calibration, high positioning accuracy and reliability, etc., and the actual measurement accuracy under dynamic conditions is better than 0.3 m in the flight direction, 2.5 cm in elevation direction and 5 cm in yaw direction obtained from midline video, meeting the test requirements.

To solve the problem of particle lack of diversity in particle filter resampling track algorithm, a particle filter algorithm based on adaptive sampling methods is presented. First, the particle weight classification, medium weight particles remain unchanged, large and small particles weights by optimized combination of weights, followed a small part of the optimization particle by system resampling. Finally, the copy number of particles in the right amount and particle compare operations. The test results show that the improved particle filter algorithm not only improves the operational efficiency of maneuvering target tracking, but also improves the stability of tracking, targeting more accurate.

The classical mean-shift tracking algorithm with the fixed bandwidth easily leads to some problems, such as the missing of the target and the mismatching of the tracking window and the target. A new mean-shift tracking method to adjust the tracking window based on partition is proposed. After the accurate location of the target centroid, the variation of the target’s scale can be reflected by the variation of the distance between the target centroid and the block centroids, which can be used to adjust the tracking window. The proposed approach is illustrated by some tracking examples with valid results.

Due to the limitations of the traditional non-parametric transform measures, a stereo matching algorithm based on non-parametric transform measure with local texture weighted item and semi-global matching method to aggregate cost is proposed. According to the directivity of the image texture metric, a contrast value of local texture is added to calculate the grayscale mean of all of the pixels in the window. The mean and the local texture contrast value are weighted sum as new matching primitive. The matching cost is determined by using semi-global matching from 8 directions. It is subsequently optimized by minimum cost to gain initial disparity. Finally, the parallax histogram of each divided region is obtained through image segmentation based on mean-shift. Peak is selected as the final disparity of each region to obtain the dense disparity map. Experimental results show that the algorithm gets more accurate results than lots of the local algorithms. It is a good solution to the distortion problem and be well adapted to the measurement of the real scene.

Panorama vision is one of crucial technologies in computer vision, and the selection of projection model plays an important role in panorama. In order to decrease the horizontal line distortion which cylinder projection brings and make conjunction between different camera-images more smooth, a multi-plane projection technology based on cylinder projection is proposed for multi-camera panorama system. Firstly, the cylinder panorama image is constructed by cylinder projection. Then cylinder panorama is divided into several sub-images horizontally by image content. At last, all sub-images are projected by anti-cylinder projection and the horizontal line distortion in the final panorama is decreased. Experiments show that the projection method can improve the horizontal line curvature to no more than 0.002, meantime ensure the smoothness between two different image planes.

An error concealment algorithm in stereoscopic video transmission based on structural similarity (SSIM) is proposed. Firstly, the concepts of temporal and inter-view SSIM are presented. Secondly, according to the temporal correlation of video sequence, the prediction mode of a Macroblock (MB) in the right view on the previous time is taken as the prediction mode of the lost MB. Thirdly, temporal and inter-view correspondence is adopted to obtain the pixel-wise SSIM map of the right view on the previous time. Finally, the prediction of the MB is obtained by comparing the SSIM values of each MB between the temporal and inter-view of the right view in the previous stereo image pairs. Consequently, the given MB prediction mode is taken as the MB mode in the lost frame, and the motion and disparity compensated prediction are used to restore the content of the lost frame. Experimental results show that the proposed algorithm has improved the PSNR by 2.76 dB and 3.43 dB compared with the traditional methods and Pang’s algorithm, respectively. Meanwhile, the proposed algorithm is efficient in improving the subjective quality of the concealed lost frames.

For the problem of low-resolution of infrared nephogram, a method of infrared nephogram super-resolution using coupled over-completed dictionary learning is presented. Based on the analysis of infrared nephogram degradation model, a super-resolution reconstruction framework is built with the sparse representation theory. First, consist a training sample by random sampling a large number of high-resolution and low-resolution sample nephogram patches. Then, construct a couple of dictionaries Dh and Dl by dictionary training, to ensure that the corresponding high-resolution and low-resolution nephogram patches have a similar sparse representation on their dictionaries. We propose a coupled dictionary training method to change the strategy of dictionary update, and the coupled dictionaries were obtained by alternatively optimizing the Dh and Dl in each step of iterative. Finally, for the inputting low-resolution infrared nephogram, Optimized Orthogonal Matching Pursuit (OOMP) algorithm is used to achieve the high-resolution infrared nephogram which satisfies the reconstruction constraint. Numerous experiments show that the proposed algorithm can get a higher quality reconstructed infrared nephogram. Moreover, the algorithm efficiency is more effective than the other dictionary-learning algorithms.

Recently, a scheme for null space based linear discriminant analysis using random matrix multiplication (NSLDA/RMM) is proposed. The computational complexity of NSLDA/RMM is still relatively high since it need compute the eigenvalue decomposition of an n×n matrix, where n is the number of the training samples. To improve the efficiency of null space based linear discriminant analysis further, we present a new and fast scheme for null space based linear discriminant analysis using random matrix multiplication (FNSLDA/RMM). FNSLDA/RMM need not compute the eigenvalue decomposition of an n×n matrix. Then the computational complexity of FNSLDA/RMM is much lower than those of the existing schemes for null space based linear discriminant analysis. Theoretical analysis of computational complexity and experiments on face database demonstrate that FNSLDA/RMM is much more efficient than the existing schemes for null space based linear discriminant analysis, but the recognition rates of FNSLDA/RMM and the existing schemes for null space based linear discriminant analysis are the same.

In order to overcome the noise sensitivity disadvantages of traditional Local Binary Pattern (LBP), this paper presents using Gauss filters to preprocess face images to remove interference noise. In order to overcome the no non-local feature extraction disadvantages of traditional LBP, this paper presents a new Multiple Weight Local Binary Pattern (MWLBP) operator. MWLBP operator weights sum spatial region histogram value of different sizes square type neighborhood regional LBP. Compared with traditional LBP, this new operator extracts features in a much larger area and can preserve a certain non-local features while extracting local features at the same time. Compared with Gabor feature and other feature extraction methods, this new operator can control the amount of calculation while preserving multiple scale features. Numerical experimental results in ORL and Yale face datasets show that Gauss filter preprocessing can remove interference noise, and improve recognition accuracy rate. MWLBP has smaller computational complexity, less classifier training time, faster operational efficiency and higher recognition accuracy rate than traditional LBP, Gabor feature and other feature extraction methods.

The utility problem will occur after the case-based reasoning system runs many times, and this problem results in a decrease performance, such as a large storage space, a low retrieval rapid. To solve this problem, a cases reduction algorithm based on support vector is proposed, whose core idea is to study the distribution principle of all cases, and find support vectors and then decide 3 boundaries, at last, reduce the redundant cases which are not of use to solve new problem. Experimental tests indicate that, the method reduces cases in the case-base and decrease retrieval complexity at the expense of very small classification accuracy.