Applied Mechanics and Materials Vols. 433-435

Abstract: Embedded environmental vision is a key issue for robotics. However, the image data is large, which usually will seriously affect the system processing speed and performance. Aiming at the feasibility and the real-time performance of robotic embedded vision system, by combining the up-to-date compressed sensing technology, a novel wavelet sparsity based simple deterministic 0-1 measurement matrix (0-1SDMM) is designed. The simulation results in matlab environment show that the 0-1SDMM has better performance than traditional Gaussian matrix in reconstruction result and reconstruction time. It provides an important reference for the future robotic embedded vision system.

Abstract: Human body segmentation is important for object tracking and recognition. When there are multiple human bodies, because of inter-occlusion, human body precise segmentation is difficult. A segmentation method based on prior shape model and level set is proposed. Human coarse shape models are constructed with position, scale and posture. For each human body, its corresponding human shape model is obtained by model matching by which position is obtained roughly after model matching, and object precise contour is obtained through curve evolution by multiphase level set with initial contour obtained from shape model. The proposed method could segment human object precisely.

Abstract: This paper proposed a lane detection algorithm for urban environment. The algorithm was concerned on selecting an appropriate limited region of interest (ROI) by OTSU segmentation. Then candidates of lane markers were extracted by Canny, finally the lane boundaries were detected by Hough transform. The limited ROI helps to identification lane in an appropriate region. This process have the effect of enhancement in the speed of operation. The proposed algorithm was simulated in MATLAB. The test databases were shared by Fondazione Bruno Kessler (FBK). The experiments show that lane boundaries can be detected correctly although they are fade. Feature-based method is usually affected by intension of image. Several characteristics of roads need to be considered further for detection more precisely.

Abstract: In computer vision system, the corners position need to be extracted from plane plate image. This paper presented a novel algorithm that improved the accuracy of corner detection from pixel to sub-pixel. The Canny operator was used to detect the corner edge pixels, and Gaussian filter was substituted by the bilateral filtering. It can remove noise and retain more slight edge information. Then, the corner edge pixels were transformed by Zernike moment and the sub-pixel edges of the corner were getted. Finally these sub-pixel edge points were linearly fitted and it was resulted in the corner coordinates of the intersection of two fitting straight lines. The experimental results show that the proposed method improves the corner detection precision to 0.1 pixel.

Abstract: A pattern recognition method based on modified perception algorithm was proposed in this paper. The principle of this method is minimizing the sum of distance between the wrong classified samples and the interface of different categories, and then the iteration was done with information contained in the minimized errors. The experiment shows that the proposed method can classify the images from Yale, ORL and actual face databases accurately and effectively. The main advantage of the method is that the algorithm is convenient for computer programming, and the recognition rate and recognition time have been distinctly developed.

Abstract: In order to solve the problem of the color characteristic information is often ignored in color image edge detection process, we propose a new color image sub-pixel edge detection method, which uses Ostu algorithm to get coarse positioning edges and extracts sub-pixel edge on the combination of curve fitting and coarse positioning edge in the projected image obtained by dimensionality reduction technique.Experimental results show that the algorithm positioning accuracy can reach 0.14 pixels,it is able to extract color image edge information effectively.

Abstract: This paper extends our previous algorithm for clustering. This previous algorithm works fine on simulated data. It can acquire satisfactory clustering results even with annular or zonal simulated data by causing the data to shrink within a cluster. To make use of the advantages of the previous algorithm, a one-dimensional (1D) histogram is mapped to a two-dimensional (2D) image and can be clustered by the previous algorithm, thus leading to stable results of histogram thresholds. The shrinking procedures of the 2D image or the 1D histogram are given, and a new parameter strategy is discussed.

Abstract: Video summaries provide a compact video representation preserving the essential activities of the original video, but the summaries may be confusing when mixing different activities together. Summaries Clustered methodology, showing similar activities simultaneously, enables to view much easier and more efficiently. However, it is very time consuming in generating summaries, especially in calculating motion distance and collision cost. To improve the efficiency of generating summaries, a parallel video synopsis generation algorithm is proposed based on GPGPU. The experiment result shows generation efficiency is improved greatly through GPU parallel computing. The acceleration radio can reach at 5.75 when data size is above 1600*960*30000.

Abstract: In image processing, removal of noise without blurring the image edges is a difficult problem. Aiming at orthogonal wavelet transform and traditional thresholds shortage, a new adaptive threshold image de-noising method which is based on wavelet packet transform and neighbor dependency is proposed. Low frequency part and high frequency part can be decomposed at the same time in wavelet packet transform and the information contained in wavelet coefficients is redundant. Using this kind of relativity in wavelet packet coefficients, we use a new variance neighbor estimation method and then neighbor dependency adaptive threshold is produced. From the experiment result, we see that compared with traditional methods, this method can not only effectively eliminate noise, but can also well keep original images information and the quality after image de-noising is very well.

Abstract: A modified compressive sensing image fusion algorithm is proposed in this paper that is based on the NSCT transform. The algorithm is improved by introducing the theory of compressive sensing into image fusion that uses the NSCT transform to make a specific image be sparse on which only the high frequency coefficient is specifically measured; The improved algorithm then process the image fusion by retrieving the maximal value of the gradient of the neighborhood average from the measured high frequency coefficient, and accordingly, maximizing the absolute value of the neighborhood variance to the low-frequency counterpart. Afterwards, the improved algorithm can reconfigure the fusion image by using the MSP reconfiguration algorithm with final deliverable of the fusion image by committing to the NSCT reverse transform. Simulation results show that the improved algorithm is superior to other hand-on algorithms both in visual effect and in objective evaluation. In the case that the storage and transmission data are limited, the algorithm comes forth better effect of image fusion that is verified to be possesses of high value in practice.