Search

In the past decade, the pedestrian detection has drawn much attention due to the significant role it plays in artificial intelligence system and vehicle assisted driving system. In order to achieve a balance between recognition rate and detection time, a novel architecture of pedestrian detection has been proposed in this paper. Firstly, we get the enhanced HOG feature (eHOG) by enhancing Histograms of Oriented Gradients feature contrast. Then, the eHOG is used as an input of XGBoost to recognize pedestrian/non-pedestrian. The architecture proposed is tested on MIT and INRIA pedestrian datasets, experimental results show that recognition rate can reach 91.13% on MIT dataset and 99.11% on INRIA dataset, and testify the effect of our architecture proposed.

Gait recognition has gained increasing attention from biometric scholars because of its applications in video surveillance. A novel approach to recognize human gait is presented based on multi-feature fusion and nearest neighbor classifier. The developed method firstly extracts 3 kinds of features from gait skeletal data by exploiting Kinect followed by calculating distances between any 2 gait sequences corresponding to these features. In what follows, weighted summation can be applied to acquire one fusion that is used in nearest neighbor classifier to recognize any gait sequence. The effectiveness of this gait recognition algorithm is verified on UPCV gait database and the experimental results show that the method has encouraging recognition rate.

The geospatial data nowadays are usually large and complex, which call for innovative and easy-to-understand visualization platforms that can reveal meaningful information. Compared with the traditional visual analysis methods, today’s visualization technologies have been able to deal with large amount of complex data. This paper reviews the technologies and applications of geospatial information visualization. Finally, the problems and challenges of geospatial visualization are expounded, and the possible future directions are discussed.

A novel approach toward pedestrian detection applied to infrared images using improved Shape Context feature (ISC) is proposed. We apply Bilateral Filtering for images to preserve the information of large sharp edges and to facilitate the task of Shape Context feature (SC) extracting. Then, we study an image descriptor based on an improved Shape Context feature (ISC), which is more robust to object deformation. For a test image, improved SC features are extracted and matched to the codebook. A voting scheme then obtains object locations from the matching results. Experimental results with Infrared Image Database demonstrate that our method is suitable for infrared images, and is promising.

The recognition of facial complexion plays an important role in facial diagnosis, which is an important part of inspection in Traditional Chinese Medicine (TCM). In this paper, we proposed intelligent techniques for recognition of facial complexion, which is classified into six classes, including normal, cyan, red, yellow, black, and white, based on the traditional theories of TCM. Quantification color histogram was employed to extract features from skin blocks of facial complexion images in RGB color space. These features were then converted to one-dimensional features, after that these were used to construct a classification model for recognition of facial complexion by supervised latent Dirichlet allocation (sLDA), which provides useful descriptive statistics for a collection, which facilitates tasks like browsing, searching, assessing document similarity, and classifying. We made experiments versus LDA feature based SVM. The experimental results showed that our method exhibited good average accuracy.

an approach toward pedestrian detection applied to natural images using improved Random Forest (RF) is proposed. We take a more discriminative method for object part detection by applying the feature of pixel-based. We firstly train a pedestrian random forest which directly maps the image patch appearance to the probabilistic vote about the possible location of the object centroid. For a testing image from the TUD dataset, our system requires four operations, which are feature extracting, passing patches through the trees, casting the votes, and processing the Hough images. Experimental results with the challenging TUD Image database demonstrate that the accuracy and robustness of our algorithm are better than those of ISM-based detection method, and this method is promising for object detection significantly.

The recognition of the teeth marks on tongue pays an important role for automatic tongue diagnosis in Traditional Chinese Medicine. If the teeth marks are not obvious, it is difficult to recognize the teeth marks on tongue desirably with some popular methods directly. In order to overcome this difficulty, a method that bases improved level set method, curvature and image gradient is presented. First, the tongue boundary line is initialized in the HSV color space and a method which enhances the contrast between tongue and other parts of the tongue image is introduced. Second, a new region-based signed pressure force function is proposed, which can efficiently stop the line at weak edges. Third, we use a Gaussian filtering process to further regularize the level set function instead of reinitializing signed distance function. Finally, we propose a method which recognize the teeth marks with both line curvature and image gradient. Experiments by numerous real tongue images show desirable performances of our method.

The knowledge discovery of driver behavior in traffic accident database is implemented based on multi-agent technology. The multi-agent technology helps to solve the complexity problems of both discovery and driver's behavior, and achieve more intelligence for driver behavior knowledge discovery. The application of pair-database-coordination mechanism and heuristic algorithm makes the knowledge discovery go in the direction of the system's automation convergence, and it improves the efficiency of whole discovery process. Multithread technology and different KDD (knowledge discovery in database) methods are also introduced to knowledge discovery of driver behavior in traffic accident database. The mining result shows that KDD in the database of traffic accident based on multi-agent is effective. The beneficial knowledge of driver behavior is obtained, and a framework which apply driver behavior knowledge to driving status warning system is prompted.

Tongue diagnosis is a method to analyze disease by observing tongue, which composed of observing tongue proper and tongue coating. The separation of tongue coating and tongue proper is a premise to establish a system of automatic diagnosis by the feature of tongue image in Traditional Chinese Medicine, whose qualities affect on the performance of tongue diagnosis. In this paper, we present a novel separation framework of Tongue Coating and Tongue Proper: an automatic initialization of contour curve by the feature of tongue in HSV color space. Meanwhile, an improved level set algorithm is designed to assist the separation of tongue coating and tongue proper, in which we explicitly define an external energy that can move the zero level curve toward the object boundaries. Applying our method to the large database of tongue images, we achieve promising experimental results.

Tongue diagnosis is an important foundation of the syndrome difference. The segmentation of the body of tongue is a premise to establishing a system of automatic diagnosis by the feature of tongue in Traditional Chinese Medicine (TCM), whose qualities affect on the performance of tongue diagnosis. Because of similar color between tongue body and background, it is difficult to segment tongue. In order to overcome two key difficulties with the initialization and boundary concavities (important to the analysis of tooth's marks), a novel method for tongue contour extraction based on improved level set curve evolution is proposed. We present an automatic initialization of contour by the feature of tongue in the HSV color space. Improved level set method takes not only color information but also tongue contour shape constraint represented by energy function between the evolving curve and parametric shape model. Applying our method to the large database of tongue images, we achieve promising experimental results.

in clinical Magnetic Resonance Imaging (MRI), any reduction in scan time offers a number of potential benefits ranging from high-temporal-rate observation of physiological processes to improvements in patient comfort. In this paper we proposed a reconstruction algorithm by applying contourlet thresholding in inverse scale space flows. We improved the inverse scale space with the noise item in which there is some meaning information and take full advantage of coutourlet transform's characters: perfect reconstruction, noise restraint and good directional selectivity. Several example MRI reconstructions from highly undersampled K-space data are presented for recovery of detailed features from incomplete and inaccurate measurements. Keywords-MRI; Inverse Scale Space; reconstruction;contourlet

Tongue diagnosis is one of the most valuable and widely used diagnostic methods in traditional Chinese medicine (TCM). A tooth-marked tongue is an important objective index of the diagnosis of qi (yang) deficiency syndrome during the diagnosis of the TCM. In order to accurately analyze it, an effective method to confirm the amount of tooth-marked and to attain the degree of tooth-marked is brought forward. It sufficiently applies the characters of GVF snake and the features of the curvatures and gradients in the all points of tongue contour. The experiment results demonstrate the efficiency of our proposed algorithm.

A novel method for face recognition (FR) based on a combination of global features extracted by nonlinear kernel principal component analysis and local features derived by applying Gabor wavelets is discussed. It is well known that the distribution of face images is highly nonlinear under a large variation in viewpoints. Therefore, linear methods such as principle component analysis (PCA) or linear discriminant analysis (LDA) cannot provide reliable and robust solutions for FR problems. In our framework, the proposed LDA in the unitary space makes use of the null space of the within-class scatter matrix effectively, and complex vectors integrate global feature vectors and local feature vectors as input feature of the proposed LDA. The experiment results demonstrate that the proposed methodology is more effective and robust for face recognition with complex face variations.

A personalized feature framework based on improved linear discriminant analysis (LDA) is proposed for face recognition (FR). In traditional LDA, the definition of the between-class matrix might cause large overlaps of neighbouring classes, so we add a weighting function into the between-class scatter matrix. In the framework, the improved LDA makes use of the null space the within-class scatter matrix effectively, and global feature vectors and local feature vectors are integrated by complex vectors as input features of improved LDA. The proposed method is compared to other commonly used FR methods on two face databases. Results demonstrate that the proposed method outperforms traditional FR approaches.

A combined personalized feature framework is proposed for face recognition. In this framework, the novel linear discriminant analysis makes use of the null space of the within-class scatter matrix effectively, and in order to simulate the recognition of the human visual system, global feature vectors and local feature vectors are integrated by complex vectors as input feature of linear discriminant analysis. The proposed method has been tested, in terms of classification error rate performance, on the multi-view UMIST face database. Results indicate that the proposed method is able to achieve excellent performance with only a small set of features.

A novel LDA method for tongue images recognition (TIR) based on a combination of global features extracted by nonlinear kernel principal component analysis and local features derived by applying Gabor wavelets is discussed. It is well known that the distribution of tongue images is highly nonlinear under a large variation in viewpoints. Therefore, linear methods such as principle component analysis or linear discriminant analysis cannot provide reliable and robust solutions for TIR problems. In our framework, the improved LDA in the unitary space makes use of the null space of the within-class scatter matrix effectively, and global feature vectors and local feature vectors are integrated by complex vectors as input feature of the improved LDA. The experiment results demonstrate that the proposed methodology is more effective and robust for tongue images recognition

The segmentation of the body of tongue is a premise to establishing a system of automatic diagnosis by the features of the tongue in traditional Chinese medicine, whose qualities affect the performance of tongue diagnosis. In order to overcome two key difficulties with the initialization and boundary concavities, a new active contour model is introduced as a segmentation method. We present an automatic initialization of snake by the feature of tongue in the HSV space. Multi-resolution wavelet (Gabor) decomposition is applied in an active contour model generating the edge energy of the image. Experiments on real tongue images show the good performance of this method.