This PDF file contains the front matter associated with SPIE Proceedings Volume 8003, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.
Bad weather, such as fog and haze, can dramatically degrade the visibility of a scene. To solve this problem, we exploit the method of dark channel prior to recover a single haze image, which is based on a key observation - most non-sky local patches in haze-free outdoor images contain some pixels with very low intensities in at least one color channel. In this article, the dark channel prior method is implemented to remove haze from a single outdoor image, and the setting up of parameters is discussed and evaluated. The experimental results show the visibility, contrast, and detail of the scene are significantly enhanced comparing with the original ones. In addition, the analysis results of the GMG (Gray Mean Grads), LS (Laplace operator), and BN (Blur Noise) image quality assessment methods present a dramatic rise of image quality.
Spectral clustering method has been widely used in image segmentation. A key issue in spectral clustering is how to build the affinity matrix. When it is applied to color image segmentation, most of the existing methods either use Euclidean metric to define the affinity matrix, or first converting color-images into gray-level images and then use the gray-level images to construct the affinity matrix (component-wise method). However, it is known that Euclidean distances can not represent the color differences well and the component-wise method does not consider the correlation between color channels. In this paper, we propose a new method to produce the affinity matrix, in which the color images are first represented in quaternion form and then the similarities between color pixels are measured by quaternion rotation (QR) mechanism. The experimental results show the superiority of the new method.
The criteria of detection and localization are always a pair of contradiction in edge detection, e.g. Canny. Due to the optimal geometry property of the Gamma probability density function (PDF), it is introduced in this paper as a kernel function after its definition is complemented at the origin. Besides, an edge preserving parameter ε is added to make the pair of contradiction to be adjustable independently. With Gaussian kernel function substituted by the modified Gamma PDF, an improved edge detection algorithm is proposed. For a given edge detection criterion, the localization criterion of Gamma detector is better than that of Canny. The advantage has been analyzed theoretically and validated through the experiments on airborne remote sensing power line images.
Numerous edge detection methods have been proposed to detect image edges. However, these methods are not very effective in detecting edges in strong noisy images. Recent years, multiscale analysis has been introduced to the realm of image processing. As the third generation wavelet, shearlets have their own superiority. Anisotropic dilation operator and shear operator are introduced to overcome the shortcomings of traditional wavelets. Because of their sensitivity to directions, shearlets are apt to do the job of edge detection. Based on shearlets, in this paper, a new edge detection method is proposed. The main idea about this new method is combining the shearlet denoising method with the edge detecting method based on shearlets. Analyzing results show that edges are characterized as zerocrossing points in shearlet domain and can be extracted from shearlet transform coefficients by detecting zero crossing points and using boundary tracking method. Many experiments are conducted to test this novel approach and we also compare Sobel, Log and Canny operators with this new method. Experiments demonstrate that when an image existing high deviation Gaussian noise, this method are much better than ordinary edge detection operators in time domain.
This paper proposed a method that can measure confidence for cascade classifier. It has great value in fusion of multi-classifier, sorting object detection results and so on. Differs from the traditional qualitative description results though using cascade classifier, this method could give a quantitative description for the results. Confidence evaluation function is constructed to calculate the confidence for the targets and the classifiers. Several experiments have been done and the results show that false alarm rate could be greatly reduced with little time cost.
Image segmentation is an important problem in pattern recognition, computer vision and other related area, which is still a research focus. In this paper, we consider the segmentation as pixel classification scheme and introduce a manifold way to address this problem. Some local features, such as Haar, LBP and SIFT, are used to represent each pixel in the image together with the basic property of the pixel. We put these pixel features on a manifold called pixel feature manifold (PFM) obtained via manifold learning methods and classify pixels with k-NN classifier in the pixel embedding space. Experimental results on MSRC image dataset show that our PFM method can effectively segment images.
The study on snow scope monitoring based on active and passive remote sensing data: a case of HJ-1 and ALOS
In this paper, the snow scope monitoring model based on active and passive remote sensing has been proposed to quantitatively inverse snow coverage range. The proposal model and two kinds of remote sensing data such as HJ-1 CCD/IRS and ALOS/PALSAR images have been employed to extract the snow coverage area. The results show that the proposal model has a high precision and can be used in snow disaster monitoring.
Edge detection is one of the most fundamental and important research fields in computer vision and image processing. This paper proposed a novel method, which focuses on edge fragments, not on individual pixels. The stability of each edge fragment, represented by the relative change of the skeleton over a given range of threshold, was analyzed to select the stable ones. The result of detection is a bunch of stable edge fragments in a forest structure, and the synthesis of them gives a single edge image. Experiments show that it is comparable with popular detectors in edge detection tasks, and it has got many particular advantages, like its stableness and fragment structure.
Tropical Cyclone (TC) center locating is a crucial step in analyzing TCs. TCs can be divided into Eye TC and No-eye TC. Different TCs have different characters. Before locating its center, TC should be detected from the whole satellite cloud image. Both infrared (IR) images and visible (VIS) images are used to detect TC. Some pattern recognition and image processing methods are applied in the paper. They are well used both in detecting TCs and getting the centers of Eye TCs. Gray Model and Chan-Vese model play important roles in locating the No-eye TCs' centers. Gray Model is used to predict the initial position of the next TC center. Since the cloud wall near the No-eye TC center is homocentric circle, Chan-Vese model (C-V model) is used to get TC contour. Experiments on MAN-YI (in 2007) and WUTIP (in 2007) show that the average error is within 0.2 degrees compared with the best track data supported by China Meteorological Administration.
Convective storms are dangerous atmosphere hazards often accompanied by heavy rain and strong wind. Despite their short life time and small spatial scale, Doppler weather radars can provide a 3-D high temporal and spatial resolution of continuous measurement. Thus, a combination of identification, tracking and nowcasting of convective storms based on radar data is one of the most important meteorological methodologies. However, most forecasting algorithms only use extrapolating techniques and do not reflect storm shape changes, making forecasting result unreliable. Thus, this paper presents a new method, namely level set, to forecast storms with storm cell shape changes concerned. As a result, forecasting precision is improved.
A novel automatic pavement crack detection approach based on texture feature is proposed. The bidirectional multi-level median filter is applied in pretreatment process to eliminate noise while maintain the details of crack edge. Improved center-symmetric local binary pattern (ICS-LBP) texture feature, local correlation texture feature and relative standard deviation texture feature are combined to detect the pavement cracks. Trained-decision strategy is applied to allocate each weight of features and texture features are extracted to train the weights. Experimental results show that the proposed algorithm provides better detection result in comparison with various crack extraction algorithms, and can detect the pavement crack quickly and effectively.
The characters at the billet ends are connected and broken due to the complex scene. Good segmentation of these connected and broken characters is very important in the characters recognition. However, it's difficult for the existing segmentation algorithm to deal with these connected and broken characters. A segmentation algorithm based on multiple feature decision is proposed to divide the connected and broken billet characters. It operates by projection to divide the characters roughly and then multiple feature decision to judge and divide the connected and broken characters. A series of segmentation experimental results show that the algorithm proposed is more effective than traditional segmentation algorithm in deal with the connected and broken characters.
To overcome the shortcoming in traditional edge detection, such as the losing of weak edges and the too rough detected edges, a new edge detection method is proposed in this paper. The new algorithm is based on the Nilpotent minimum fusion. First of all, based on the space fuzzy relation of weak edges, the algorithm makes decision fusion to improve the structure of weak edges by using the Nilpotent minimum operator. Secondly, detect edges based on the fusion results. As a result, the weak edges are detected. Experiments on a variety of weak edge images show that the new algorithm can actually overcome the shortcoming in traditional edge detection, for the results are much better than traditional methods. On one hand, some of the weak edges of complex images, such as medical images, are detected. On the other hand, the edges detected by the new algorithm are thinner.
The traditional edge detection algorithms have certain noise amplificat ion, making there is a big error, so the edge detection ability is limited. In analysis of the low-frequency signal of image, wavelet analysis theory can reduce the time resolution; under high time resolution for high-frequency signal of the image, it can be concerned about the transient characteristics of the signal to reduce the frequency resolution. Because of the self-adaptive for signal, the wavelet transform can ext ract useful informat ion from the edge of an image. The wavelet transform is at various scales, wavelet transform of each scale provides certain edge informat ion, so called mult i-scale edge detection. Multi-scale edge detection is that the original signal is first polished at different scales, and then detects the mutation of the original signal by the first or second derivative of the polished signal, and the mutations are edges. The edge detection is equivalent to signal detection in different frequency bands after wavelet decomposition. This article is use of this algorithm which takes into account both details and profile of image to detect the mutation of the signal at different scales, provided necessary edge information for image analysis, target recognition and machine visual, and achieved good results.
This paper proposes a novel directional asymmetric sampling search (DASS) algorithm for video compression. Making full use of the error information (block distortions) of the search patterns, eight different direction search patterns are designed for various situations. The strategy of local sampling search is employed for the search of big-motion vector. In order to further speed up the search, early termination strategy is adopted in procedure of DASS. Compared to conventional fast algorithms, the proposed method has the most satisfactory PSNR values for all test sequences.
The accuracy of target acquisition depends on preparation of templates and selection of matching measures. Consequently, the fusion of different information supplied by various metrics has important value to increase the success rate of target acquisition. The fusion of two measures is described to illustrate this method. We may firstly use template matching in terms of the first measure to find the locations and heights of top N peaks, and then compute the value under the other measure on the position of each peak. Regarding all of the peaks as the recognition frame and measures as different evidence, Dempster Combination Rules can be used to fuse the data. Furthermore, dual measures fusion can be extended to application of multiple measures. When more than two measures are employed, weights of different measures are unnecessary to be assigned artificially but gain from the distances between every two pieces of evidence. Some typical targets of urban tall buildings are used to test the performance of template matching with measures fusion. The experimental data validates the fusion of multi-measures is effective to improve the capability of target recognition.
In this paper, a method to enhance the fingerprint image by using Log-Gabor filters is proposed. Firstly, a filter for extracting fingerprint image texture feature is designed. Then, the high frequency components of fingerprint image are extracted by filtering. Finally, the fingerprint image details can be improved by enhancing high frequency components. Experimental results show that the proposed algorithm can effectively improve the quality of fingerprint image and the reliability of fingerprint identification.
A new solution to recover 6DoF (Degrees of Freedom) ego-motion is present. The problem is to estimate the ego-motion information solely from dense optical flow (OF) field efficiently and robustly, free of Inertial Measurement Unit (IMU). The algorithm is a hierarchical framework and in each level there exist three parts, which are the optical flow Computation(OFC), the ego-motion estimation(EME) from two different models, and image warping(IW) according to the EME for the next level. The numerical precision of the algorithm under noise was investigated in the paper. We also compared its performance with Srinivasan's interpolation method and the 4DoF affine model on real aerial images. Our method are more accurate under large displacements and can resist the impacts of the rotations around x and y axis in a reasonable extent during computer navigation simulation.
In this paper, a simple and flexible framework based on the duality of obstacle detection and ground homograph estimation is proposed for obstacle detecting during autonomous navigation. The virtue of which is that the obstacle can be detected and localized without requiring the parameters of the camera. The iterative frame is implemented in our method to avoid the influence of the obstacles when estimating the ground homography. We initially estimate a homograph for the ground using matching points to register the image pairs; then, the two images are differenced so that the homograph disparity image (HDI) can be generated; after that, we segment the HDI to obtain a coarse location of obstacles; and the matching points out of the obstacle regions will be considered as input for the latter loop. This process is iterated until the results converge. The feasibility of this scheme is analyzed, and the experimental results demonstrate it can gain robust results for obstacle detection.
To detect IWST, foreign and domestic scholars put forward many meaningful detection methods. However, most of the algorithms are too much complex in the calculation to meet real-time and reliability requirements in practical application. A simple self-adaptive threshold algorithm to capture and track the IWST is presented in this paper. Testing results showed that algorithm not only effectively extract IWST and track it through the clouds in the sky background, but also has a strong robustness for the interference of background noises.
In order to reduce the bad influence of undulate background in dim infrared targets detection, a new filtering method based on anisotropy is presented. Anisotropic partial differential equations have strong ability to remain direction characteristics, and it can perfectly restrain stable background and preserve undulate background. But we need simultaneously to restrain stable and undulate background, so a few improvements are done in the following. Firstly, we improve the edge stopping function of anisotropic differential, which make it increase following the value of direction grads. Secondly, we choose the two least direction values of edge stopping function as parameter values based on the direction characteristics of grads operator. For that, we are able to simultaneously restrain stable and undulate background, and deal with different characteristic regions respectively. The simulation experiments prove the effectiveness of the proposed algorithm, which detection ability is much better than isotropic filtering methods especially for applying in infrared images with complex background.
An imaging modeling simulation research of infrared aero-optical effect based on superposition of Gaussian Mixture Model
By the influence of thermal effect and incoming flow transmission effect from the surface of head-covering, the highspeed aircraft equipped with optical imaging guidance system, when flying in the dense and high-altitude atmosphere, will cause the rapid degradation of the imaging quality. To study the influence of image diffusion, image drift and image dithering caused by optical transmission effect, this paper presents a macroscopic method of simulating the degrade image by adopting superposition of Gaussian mixture model, and combining the model with the analysis of flow field and a variety of flight parameters of high-speed aircraft, Experimental results show that the simulation image is consistent with that of the theoretical analysis.
Robust lane detection and tracking approach using improved Hough transform and Gaussian Mixture Model is proposed in this paper. The approach consists of three parts: lane markings detection, lane parameters estimation and lane position tracking. Firstly, lane marking pixels are extracted using edge and color features. Then, these pixels are used to estimate the lane boundaries. After the vanishing point has been predicted by a RANSAC algorithm, we use an improved Hough transform to detect the straight lane boundaries in the near field, and apply a parabolic model to represent curved lanes probably existed in the far field. Finally, a novel lane parameters determination method, which uses Gaussian Mixture Model to represent and update the parameters of lane boundaries, is proposed to ensure the stability of the lane tracking system. The proposed approach is tested with some real videos captured on a highway with challenging road environments, and the results demonstrate that our system is very reliable and can also be implemented in real-time.
A matching-unscented Kalman filtering for gravity aided navigation is presented in this paper. With this method submerged position fixes for autonomous underwater vehicle can be obtained from comparing gravity fields' measurements with gravity maps, meanwhile the drawback of traditional matching or filtering algorithms can be avoided. A synthetic gravity map was taken for the simulation, and the results showed that navigation errors can be reduced more efficiently and reliably by the presented method.
A human-machine cooperative path planning model based on cloud model is proposed in this paper. The system enables the planner take part in the A* searching process and the cloud model integrates fuzziness with randomness of the qualitative concept. In the process of human-computer cooperation, the position of the leading field is figured out based on cloud model; it effectively guides the A* searching process and avoids the drawback of the algorithm. Experiment results demonstrated the validity and the feasibility of the model. It's much more efficient than either a human or a computer algorithm in the path planning tasks.
Rock is one kind of dangerous hazards for lunar landing, a small rock about 0.5-1.0 meter tall can damage the lander and make the whole mission failed1. In this paper, we provide an improved method for rock detection based on shadows and texture for safe lunar landing, which can detect and model the rocks in the lunar surface image precisely. Although traditional rock detection algorithms can get a good result in the experimental environment full of rocks, it does not work well in the real, uneven environment with all kinds of hazards. Our algorithm can adapt to various environment because of the precisely extracting shadows of rocks and texture analyses after rock modeling.
This paper analyzes the difference between the imaging mechanism of the infrared images and that of the visible light images, and find that it is important to extract the stable and reliable common feature for object recognition. Then we propose a target recognition algorithm based on histograms of oriented gradients (HOG) which evaluates normalized local histograms of image gradient orientations in a dense grid. Last we adopt linear SVM trained for a binary object/non-object classifier and detect the object in the forward-looking infrared (FLIR) images. The experiment results suggest that the proposed approach has high rates of detection. Furthermore, we study how to select positive and negative samples for a better performance.
Color image segmentation draws a lot of attention recently. In order to improve efficiency of spectral clustering in color image segmentation, a novel two-stage color image segmentation method is proposed. In the first stage, we use vector gradient approach to detect color image gradient information, and watershed transformation to get the pre-segmentation result. In the second stage, Nyström extension based spectral clustering is used to get the final result. To verify the proposed algorithm, it is applied to color images from the Berkeley Segmentation Dataset. Experiments show our method can bring promising results and reduce the runtime significantly.
In this paper, a simple but effective method for robot self-localization is presented. The spatial neighborhood constraint is incorporated into the preprocessing of the image segmentation. Then it uses a closed cycle with rectification and Hough detection to find the boundary and corners. Depending on the actual size of surrounding environment and the white lines and corners detected last step, the robot can maintain self-localization through two methods. One method uses the two lines, and the other method used triangulation. Finally, a weight value is set between the two methods to realize the self-localization.Actual image sequence from the robot is tested. The robot can be placed anywhere in the environment. The final self-localization results on very different images with significant light change and noise are promising.
Infrared small target detection is an important research area of computer vision and often a key technique in Infrared Search and Track (IRST) systems. Many algorithms have been reported for this purpose. The facet-based method is one of novel algorithms and is shown as robust and efficient, but it does not perform well in target preservation. The method cannot detect peripheral pixel of target, which causes information loss of target intensity distribution and affects post processing of detection, such as target tracking and recognition. In this paper an improved algorithm is developed for solving this shortcoming. The detection behavior of the facet model is further analyzed. Small target is surrounded by background, so local image edge that indicates target contour can be represented by zero-crossings of the second partial derivatives. The improved algorithm uses facet model to fit local intensity surface and detect potential targets using extremum theory, then the zero-crossings of the second partial derivatives of the fitting function in each potential target's neighborhood are found and the pixels inside the zero-crossing contour are restored to the potential target. In experiments involving typical infrared images target intensity distribution information is well preserved by proposed algorithm and its execution time is also acceptable.
This paper concerns the problem of fast vehicle license plate location. A new method has been proposed for locating the vehicle license plate in the color image with a high speed. The color image is transformed into HSV color space and each single channel image is operated by different operations such as image equalization, image binary and so on. Then the results of each channel image is integrated, and a mathematical morphology method together with image smoothing, image filtering and contour extraction are used to get the candidate vehicle license plate. Finally, a minimum rectangle enclosing the extracted contour is obtained, and affine transformation is done to the rectangle. The experimental results of more than 140 images collected demonstrated that the accuracy of detection is 95.21%, and the average time cost for each image is about 50 milliseconds.
Keywords: Unmanned Aerial Vehicles (UAVs) are been increasingly used in civilian and military domains. Vision-aided inertial navigation system in UAV is studied by more and more researchers for it's non-contact, high accuracy and stability. In this paper, an L1-Graph-based image matching approach, which constructs neighbouring system based on sparse representation, is proposed for monocular motion vision measurement. Then, a scheme for amending the outputs of inertial sensor for the velocity measurement is designed, which fuse the outputs from the downward-looking velocity measurement and inertial sensor by Kalman filter. Experiments show this design form an accurate navigation solution.
Focusing on downward-looking infrared cloud images, we propose a cloud detection and thin clouds removal algorithm based on the feature judgement. This algorithm determines whether infrared image contains clouds or not by texture feature extraction, segments the suspected cloud regions and determines the cloud region using a voting system based on multi-feature quantitative analysis. Finally, we enhance image and remove the cloud interference using the combination of space and frequency domain. Test results show this algorithm has good practicability and accuracy, which improves the automation and intelligence of remote sensing information processing and application.
This paper proposes a novel level set based image segmentation method by use of image second statistics and logarithmic Euclidean metric. Different from previous tensor based image segmentation approaches, the proposed method adopts covariance feature as region-level descriptor rather than pixel-level one. On the basis of feature image, we utilize second order statistics of image feature, i.e., covariance matrix, to model image region representation, which is of low dimension, invariant to uniform illumination change, insensitive to noise, and more importantly provide a natural mechanism of incorporating different types of image features by modeling their correlations. We model image segmentation problem as one finding the optimal segmentation that maximizes the covariance distance between foreground region and background region. Typically, covariance matrices do not lie on Euclidean space. Our solution to this is to exploit logarithmic Euclidean distance as a metric to compute the similarity between two matrices. The experimental results show that covariance matrix as region descriptor do form an effective representation for image segmentation problems, and the proposed image energy can be used to segment images and extract object boundaries reliably and accurately.
To tackle occlusions, a hierarchical part matching method based on a layered appearance model for object tracking is presented in this paper, which integrates global and partial region matching together to search the target object in a coarse to fine manner. In order to reduce the ambiguity of object localization, only the discriminative parts are selected for similarity computing with respect to their cornerness measure. The similarity between parts is computed in a layerwised manner, based on which the state of occlusions can be inferred correctly. When occluded partially, the object can be localized accurately, when occluded completely, the historical information of motion is applied to predict its position by a Kalman filter. The proposed tracking method is tested on practical video sequences, and the experimental results show it can consistently provides accurate positions of the object for stable tracking, even under severe occlusions.
The purpose and method of image quality assessment are quite different for automatic target recognition (ATR) and traditional application. Local invariant feature detectors, mainly including corner detectors, blob detectors and region detectors etc., are widely applied for ATR. A saliency model of feature was proposed to evaluate feasibility of ATR in this paper. The first step consisted of computing the first-order derivatives on horizontal orientation and vertical orientation, and computing DoG maps in different scales respectively. Next, saliency images of feature were built based auto-correlation matrix in different scale. Then, saliency images of feature of different scales amalgamated. Experiment were performed on a large test set, including infrared images and optical images, and the result showed that the salient regions computed by this model were consistent with real feature regions computed by mostly local invariant feature extraction algorithms.
This paper presents a method for ship target detecting in complex background. It aims at solving two difficulties in detection. The first one is that the ships docking in-shore cannot be segmented because of its gray level similarity to land, and the second is that the ships linked side by side cannot be easily located as separate correct target. The first one is solved by extracting water region firstly by measure of harbor-template matching. In order to reduce the impact of angle difference which leads to error, we update the template by the corresponding angle calculated recur to line feature. Then matching fine with the updated template to extract water region wholly in which the segment is effective. For the second difficulty, the smallest minimum bounding rectangle (SMBR) of the segmented areas are obtained by contour tracing, and the areas are projected to the two different directions of its SMBR, then the projection curves are acquired. If the ships are linking together, the peak-valley-peak pattern will exist in the projection curve and the valley-point indicates the ships' connection position. Then the ships can be separated by cutting the area at connection position along the projection direction. The experiment results verify the efficiency and accuracy of our method.
To improve the performance of traditional random walk algorithm, an image segmentation algorithm is proposed, which combined random walk and data-adaptive gaussian smoother. Because the medical or remote sensing images are often occupied by strong noises, a data-adaptive anisotropic filtering technique is proposed to remove noise, The filtering technique built on top of an iterative scheme that can preserve the original significant structures while suppressing the noises to the largest extent, and then compute the gradient image of the filtering image. At last the weights of edges of random walk are determined by both the gray value of original image and the salient features of data-adaptive gaussian smoother. The experimental results from synthetic as well as real images demonstrate that the proposed approach is more effective, accurate and more robust in the noise.
Stable imaging tracking method based on learning online for ground moving target with multi-DSP processing
A stable imaging tracking method based on learning online for ground moving target with multi-DSP processing is presented in this paper. Background window is set to track and predict the background image and supervise the intruder. The target learning online based on background prediction revises the accumulated tracking error. Different tracking strategy during different tracking states and risk level of intruder improves the stability and accuracy of tracking system especially in a long time of continual tracking. The parallel processing based on multiple DSP makes a real-time tracking system be possible.
Compared with the traditional EM clustering algorithm, the EMBoost clustering algorithm can improve two problems that the sensitive result to initial value and the low precision. However, an important factor, the local information, is not considered in the EMBoost algorithm, which is useful to enhance the performance of the EMBoost algorithm, especially for image segmentation. We believe that neighbor pixels to the center measured by the space distance and the texture distance are beneficial to the internal consistency of the homogeneous region. Hence, we proposed a new approach that spatial information is brought into EMBoost clustering algorithm, which consisted of the adjacent pixels relative position and the neighbor texture distance, in order to improve the performance EMBoost clustering method. According to the experimental results of the texture image segmentation and the Synthetic Aperture Radar (SAR) image segmentation, the proposed method can obtain better accuracy and visual effect, compared against other methods.
An automatic tracking algorithm based on Camshift and Kalman filter is proposed in this paper to deal with the problems in traditional Camshift algorithm, such as artificial orientation and increasing possibility of tracking failure under occlusion. The inter-frame difference and canny edge detection are combined to segment perfect moving object region accurately, and the center point of the region is obtained as the initial position of the object. With regard to tracking under occlusion, Kalman filter is used to predict the position and velocity of the target. Specifically, the initial iterative position of Camshift algorithm is obtained by Kalman filter in every frame, and then Camshift algorithm is utilized to track the target position. Finally, the parameters of adaptive Kalman filter are updated by the optimal position. However, when severe occlusion appears, the optimal position calculated by Camshift algorithm is inaccurate, and the Kalman filter fails to estimate the coming state effectively. In this situation, the Kalman filter is updated by the Kalman predictive value instead of the value calculated by the Camshift algorithm. The experiment results demonstrate that the proposed algorithm can detect and track the target object accurately and has better robustness to occlusion.
Infrared image quality assessment plays an important role in FLIR. In this paper, infrared image quality assessment specified in gray scale template matching was studied. Target local salient metric (LSM) and global salient metric (GSM) were presented as quality indices, and Back-Propagation network was adopted to integrate the indices. Experiment shows that error caused by the proposed method is less than the error caused by the classical indices, such as ETB and TIR.
To the problems of blur in aero degraded image due to noise disturbance and aero-optical effects, an adaptive wiener filter algorithm of aero degraded image based on precise image segmentation was proposed. Firstly, edge detection was done to precisely comminute edge region. Then, adaptive wiener filter was done on the premise that point spread function (PSF) submits to Gaussian distributing.thereinto, image definition estimate function was used to dispose controllable gene of algorithm to fulfill adaptive algorithm. Numerical experimental results of aero degraded images and contrasts with other algorithms show that this algorithm is better in detail-preserving and debluring of aero degraded image. Meantime, complexity of this algorithm is lower. So, it has better property of real time.
A novel scale and rotation invariant ship recognition method using log-polar mapping and two-dimensional principal component analysis (2DPCA) is proposed. Log-polar mapping is very useful for eliminating the rotation and scale effects. 2DPCA is used to extract ship feature from normalized log-polar image become scale and rotation invariant. Experimental results show that the proposed method has improved recognition performance.
More and more infrared imaging systems were applied in military and civil fields and the performance of detecting dim target is one important index for infrared imaging system. The performance of the method based on background suppression is declined with many false alarms and even loss target in image with clutter background. In order to improve the performance of dim target detection, the method base on target enhance is developed. In this paper, after analyzing the grey value distribution of target, background and noise in local area, the pixel's shape parameter was defined, maximum strategy was used to obtain one size of pixel's shape parameter, and enumerate strategy was used to search the possible target size. And then, one character filter based on shape parameter was designed to enhanced target and suppress clutter background. Compared with the method based on background suppression, experimental results show that proposal method is effective to enhance target and suppress clutter background.
Avoiding potential safety hazard is the primary task of vision-based assistant driving system(ADS). Potential safety hazard exists in driving individual vehicles. Although these hazards are unexpected, obvious characteristics exist for vehicles that make them happen, such as: relatively fast speed, changing lanes frequently and being occluded as shuttling in the busy traffic. All these characteristics go against on-road tracking for the unsafe vehicle. At present, the assistant driving system is only permitted in the field of obstracle detection and location. However, those systems are not involved in tracking of vehicles with potential safety hazard. The paper presents an approach to tracking and online learning of on-road vehicles with potential safety hazard. Further, we improve the method of online learning to the unsafe hazard. The performance of our tracking algorithm is evaluated on a public benchmark with test data from various challenging videos on different conditions. The experiment results demonstrate that, in the same condition, our method can obtain samples more efficiently and lead the classifier to converge more quickly.
Concerning the target losing problem caused by partial occlusion, uneven illumination and so on during the image tracking procedure by traditional algorithms, a new tracking method based on robust similarity measurement criteria is proposed in this paper. The measurement is carried out with the original gray image pair with no preprocessing such as feature extraction. First, a two-dimensional joint distribution histogram is set up to describe the gray value change between the pixel in the reference and its corresponding pixel in the real image. Then, the law of data distribution under uneven illumination and partial occlusion is concluded. At last, a strategy based on Hough transformation is adopted to calculate the measurement value on the joint histogram. The method is robust for the uneven illumination. And small partial occlusion has little influence on the output value of the similarity measurement. Comparing experimental results show the new method's efficiency.
This paper proposes an online learning boosting method based on kernel regression for robust visual tracking. Although much progress has been made in using boosting for tracking, it remains a big challenge to get a robust tracker that is insensitive to illumination change, clutter, object deformation, and occlusion. In this paper, we use a nonlinear version of the recursive least square (RLS) algorithm so as to derive weak classifiers for visual tracking, which performs linear regression in a high-dimensional feature space induced by a Mercer kernel. In order to alleviate the computational burden and increase efficiency, we apply online sparsification to filter samples in feature space. In our boosting framework, adaptive linear weak classifiers are performed, the form of which is modified adaptively to cope with scene changes in every frame. Experimental results demonstrate that our proposed method has advantages in dealing with complex background in visual tracking, and often outperforms the state of the art on the popular datasets.