This PDF file contains the front matter associated with SPIE-IS&T Proceedings Volume 6813, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

We propose a new concept, called "real world crawling", in which intelligent mobile sensors completely recognize environments by actively gathering information in those environments and integrating that information on the basis of location. First we locate objects by widely and roughly scanning the entire environment with these mobile sensors, and we check the objects in detail by moving the sensors to find out exactly what and where they are. We focused on the automation of inventory counting with barcodes as an application of our concept. We developed "a barcode reading robot" which autonomously moved in a warehouse. It located and read barcode ID tags using a camera and a barcode reader while moving. However, motion blurs caused by the robot's translational motion made it difficult to recognize the barcodes. Because of the high computational cost of image deblurring software, we used the pan rotation of the camera to reduce these blurs. We derived the appropriate pan rotation velocity from the robot's translational velocity and from the distance to the surfaces of barcoded boxes. We verified the effectiveness of our method in an experimental test.

In this paper we propose a general, object-oriented software architecture for model-based visual tracking. The library is general purpose with respect to object model, estimated pose parameters, visual modalities employed, number of cameras and objects, and tracking methodology. The base class structure provides the necessary building blocks for implementing a wide variety of both known and novel tracking systems, integrating different visual modalities, like as color, motion, edge maps etc., in a multi-level fashion, ranging from pixel-level segmentation, up to local features matching and maximum-likelihood object pose estimation. The proposed structure allows integrating known data association algorithms for simultaneous, multiple object tracking tasks, as well as data fusion techniques for robust, multi-sensor tracking; within these contexts, parallelization of each tracking algorithm can as well be easily accomplished. Application of the proposed architecture is demonstrated through the definition and practical implementation of several tasks, all specified in terms of a self-contained description language.

Object tracking is an essential problem in the field of video and image processing. Although tracking algorithms working on gray video are convenient in actual applications, they are more difficult to be developed than those using color features, since less information is taken into account. Few researches have been dedicated to tracking object using edge information. In this paper, we proposed a novel video tracking algorithm based on edge information for gray videos. This method adopts the combination of a condensation particle filter and an improved chamfer matching. The improved chamfer matching is rotation invariant and capable of estimating the shift between an observed image patch and a template by an orientation distance transform. A modified discriminative likelihood measurement method that focuses on the difference is adopted. These values are normalized and used as the weights of particles which predict and track the object. Experiment results show that our modifications to chamfer matching improve its performance in video tracking problem. And the algorithm is stable, robust, and can effectively handle rotation distortion. Further work can be done on updating the template to adapt to significant viewpoint and scale changes of the appearance of the object during the tracking process.

Compact imaging devices are desirable in many different machine vision applications. For instance, in inspection for semiconductor manufacturing systems, the reduction in feature size demands lenses with a small working distance but a wide field of view. An integrated computational imaging system has proved to be advantageous in this respect, as it integrates the optics, optoelectronics, and signal processing together in the system architecture. This allows for unconventional optical systems that require further image processing to reconstruct the images, but can be made to satisfy more stringent design constraints such as size, power, and cost. In this paper, we focus on a multi-lens optical architecture. We explain the possible designs and discuss the reconstruction of images, as we need to combine the multiple low-resolution images formed from the different optical paths into a high-resolution image. We will also explore its applicability in various machine vision applications.

In the context of sustainable agriculture, matching accurately herbicides and weeds is an important task. The site specific spraying requires a preliminary diagnostic depending on the plant species identification and localisation. In order to distinguish between weeds species or to discriminate between weeds and soil from their spectral properties, we investigate a spectral approach developing a catadioptric bi-spectral imaging system as a diagnostic tool. The aim of this project consists in the conception and feasibility of a vision system which captures a pair of images with a single camera by the use of two planar mirrors. Then fixing a filter on each mirror, two different spectral channels (e.g. Blue and Green) of the scene can be obtained. The optical modeling is explained to shot the same scene. A calibration based on the inverse pinhole model is required to be able to superpose the scene. The choice of interferential filters is discussed to extract agronomic information from the scene by the use of vegetation index.

A range imaging camera produces an output similar to a digital photograph, but every pixel in the image contains distance information as well as intensity. This is useful for measuring the shape, size and location of objects in a scene, hence is well suited to certain machine vision applications. Previously we demonstrated a heterodyne range imaging system operating in a relatively high resolution (512-by-512) pixels and high precision (0.4 mm best case) configuration, but with a slow measurement rate (one every 10 s). Although this high precision range imaging is useful for some applications, the low acquisition speed is limiting in many situations. The system's frame rate and length of acquisition is fully configurable in software, which means the measurement rate can be increased by compromising precision and image resolution. In this paper we demonstrate the flexibility of our range imaging system by showing examples of high precision ranging at slow acquisition speeds and video-rate ranging with reduced ranging precision and image resolution. We also show that the heterodyne approach and the use of more than four samples per beat cycle provides better linearity than the traditional homodyne quadrature detection approach. Finally, we comment on practical issues of frame rate and beat signal frequency selection.

This paper aims at reviewing the recent published works dealing with industrial applications which rely on polarization imaging. A general introduction presents the basics of polarimetry and then 2D and 3D machine vision application are presented as well as the latest evolution in term of high speed polarimetric imaging.

Focus-based depth (Z) measurements are used extensively in industrial metrology and microscopy. Typically, a peak in the focus figure-of-merit of a region is found while moving the lens towards or away from the surface, allowing local recovery of depth. These focus-based measurements are susceptible to errors caused by: (1) Optical aberrations and characteristics of the lens (astigmatism, field curvature); (2) Optical and image sensor misalignments; (3) Image sensor shape errors. Depth measurements of the same artifact can therefore significantly vary depending on the prevailing orientation of the surface texture (due to lens astigmatism) or on the specific position in the field of view. We present a vision-based algorithm to reduce errors in focus-based depth measurements. The algorithm consists of two steps: 1. Offline calibration: We generate a calibration table for the optical system, consisting of a set of Z calibration curves for different locations in the field of view. 2. Run-time correction: During measurement, we determine the Z correction to the focus position using the stored Z calibration curves and a measurement of the local orientation of the surface texture. In our tests, the correction algorithm reduced the depth measurement errors by a factor of 2, on average, for a wide range of surfaces and conditions.

As nowadays the industry aims at fast and high quality product development and manufacturing processes a modern and efficient quality inspection is essential. Compared to conventional measurement technologies, industrial computer tomography (CT) is a non-destructive technology for 3D-image data acquisition which helps to overcome their disadvantages by offering the possibility to scan complex parts with all outer and inner geometric features. In this paper new and optimized methods for 3D image processing, including innovative ways of surface reconstruction and automatic geometric feature detection of complex components, are presented, especially our work of developing smart online data processing and data handling methods, with an integrated intelligent online mesh reduction. Hereby the processing of huge and high resolution data sets is guaranteed. Besides, new approaches for surface reconstruction and segmentation based on statistical methods are demonstrated. On the extracted 3D point cloud or surface triangulation automated and precise algorithms for geometric inspection are deployed. All algorithms are applied to different real data sets generated by computer tomography in order to demonstrate the capabilities of the new tools. Since CT is an emerging technology for non-destructive testing and inspection more and more industrial application fields will use and profit from this new technology.

Geometric modeling and haptic rendering of textile has attracted significant interest over the last decade. A haptic representation is created by adding the physical properties of an object to its geometric configuration. While research has been conducted into geometric modeling of fabric, current systems require time-consuming manual recognition of textile specifications and data entry. The development of a generic approach for construction of the 3D geometric model of a woven textile is pursued in this work. The geometric model would be superimposed by a haptic model in the future work. The focus at this stage is on hand-woven textile artifacts for display in museums. A fuzzy rule based algorithm is applied to the still images of the artifacts to generate the 3D model. The derived model is exported as a 3D VRML model of the textile for visual representation and haptic rendering. An overview of the approach is provided and the developed algorithm is described. The approach is validated by applying the algorithm to different textile samples and comparing the produced models with the actual structure and pattern of the samples.

Non-negative matrix factorization of an input data matrix into a matrix of basis vectors and a matrix of encoding coefficients is a subspace representation method that has attracted attention of researches in pattern recognition in the recent period. We have explored crucial aspects of NMF on massive recognition experiments with the ORL database of faces which include intuitively clear parts constituting the whole. Using a principal changing of the learning stage structure and by formulating NMF problems for each of a priori given parts separately, we developed a novel modular NMF algorithm. Although this algorithm provides uniquely separated basis vectors which code individual face parts in accordance with the parts-based principle of the NMF methodology applied to object recognition problems, any significant improvement of recognition rates for occluded parts, predicted in several papers, was not reached. We claim that using the parts-based concept in NMF as a basis for solving recognition problems with occluded objects has not been justified.

A circle detection method utilizing Radon transform is proposed. In this paper, a closed form solution for the Radon transform of a circular structure is derived from the Radon transform of a round disk. Because the Radon transform of circle has the unique property of invariance to the angle change, a universal matched filter can be constructed from the Radon transform of circle. To detect if there is a circular structure of specific radius presented in the image, a pre-defined matched filter is applied to the Radon transform of the image at all angles and a circle presence intensity image is reconstructed from the filtering results through filtered back projection (FBP). By thresholding the circle presence intensity image, the presence and the location of the circle can be easily determined. The preliminary experimental results show that the proposed method is effective and has better signal to noise ratio in output compared to the typical Hough transform approach.

Registration of maps and airborne or satellite images is an important problem for tasks such as map updating and change detection. This is a difficult problem because map features such as roads and buildings may be mis-located and features extracted from images may not correspond to map features. Nonetheless, it is possible to obtain a general global registration of maps and images by applying statistical techniques to map and image features. Finer analysis can then be used to find changes and local mismatches. The Maximization of Mutual Information (MMI) technique has proven to be very robust in image-to-image registration. This paper extends the MMI technique to the map-to-image registration problem through a focus-of-attention mechanism that forces MMI to utilize correspondences that have a high probability of being information rich. The number of registration parameters can be adjusted to meet the characteristics of the matching problem and accuracy requirements of the application. Experimental results demonstrate the robustness and efficiency of the algorithm.

This paper presents a new approach to the segmentation of the microscopic nuclei images. First, for segmentation of the cell nuclei from background, the adaptive local thresholding is used. A threshold for adaptive local thresholding is estimated by using the gaussian mixture model and maximizing the likelihood function of gray value of cell images. After nuclei segmentation, overlapped nuclei and isolated nuclei need to be classified for exact nuclei separation. For nuclei classification, this paper extracted the morphological features of the nuclei such as compactness, smoothness and moments from training data. For overlapped nuclei classification, this paper uses a Bayesian network with three probability density functions for evidence at each node. The probability density functions for each node are modeled using the three morphological features. After nuclei classification, segmenting of overlapped nuclei into isolated nuclei is necessary. Since watershed algorithm has the problem of over-segmentation, we find makers from each overlapped nuclei and apply watershed algorithm with the proposed merging algorithm. The experimental results using microscopic nuclei images show that our system can indeed improve segmentation performance compared to previous researches, because we performed nuclei classification before separating overlapped nuclei.

Local feature-based matching methods have witnessed great success in the context of multiple view matching, object recognition and video content analysis. Naturally, one would like to (1) investigate the merits and shortcomings of feature-based approaches; and (2) to extend such approaches to general object classes matching problems. The present paper illustrates our research attempts along this direction. The proposed feature-based method is empirically justified, and demonstrates excellent robustness against intra-class variation, structure variation, scale change, rotation and background clutter.

Line fitting and moments are two different problems and most articles discuss these two problems separately. In this paper, using the constraint optimization, we relate the line fitting to moments. We show that the eigen vectors of the second order central moments are fitted line directional vectors, and the eigen value is the fitting error. Then, we further show that the line fitting errors can be computed directly from the first and second moment invariants. From the relation between line fitting and moments, we propose a mask-size independent approach to implement the line fitting for curves or object contours. The computational cost of the new approach is independent of the mask size. It is computationally efficient if compared to the conventional approach whose computational cost is proportional to the fitting mask size.

Traditional video scene analysis depends on accurate background modeling techniques to segment objects of interest. Multimodal background models such as Mixture of Gaussian (MOG) and Multimodal Mean (MM) are capable of handling dynamic scene elements and incorporating new objects into the background. Due to the adaptive nature of these techniques, new pixels have to be observed consistently over time before they can be incorporated into the background. However, pixels in the boundary between two colors tend to fluctuate more, creating false positive pixels that result in less accurate foreground segmentation. To correct this, a simple and computationally efficient edge detection based algorithm is proposed. On average, approximately 70 percent of these false positives can be eliminated with little computational overhead.

Fog and other poor visibility conditions hamper the visibility of runway surfaces and any obstacles present on the runway, potentially creating a situation where a pilot may not be able to safely land the aircraft. Assisting the pilot to land the aircraft safely in such conditions is an active area of research. We are investigating a method that combines non-linear image enhancement with classification of runway edges to detect objects on the runway. The image is segmented into runaway and non-runway regions, and objects that are found in the runway regions are deemed to constitute potential hazards. For runway edge classification, we make use of the long, continuous edges in the image stream. This paper describes a new method for edge-detection that is robust to the imaging conditions under which we are acquiring the imagery. This edge-detection method extracts edges using a locally adaptive threshold for the detection. The proposed algorithm is evaluated qualitatively and quantitatively on different types of images, especially acquired under poor visibility conditions. Additionally the results of our new algorithm are compared with other, more conventional edge detectors.

As JPEG compression at source is ubiquitous in retinal imaging, and the block artefacts introduced are known to be of similar size to microaneurysms (an important indicator of diabetic retinopathy) it is prudent to evaluate the effect of JPEG compression on automated detection of retinal pathology. Retinal images were acquired at high quality and then compressed to various lower qualities. An automated microaneurysm detector was run on the retinal images of various qualities of JPEG compression and the ability to predict the presence of diabetic retinopathy based on the detected presence of microaneurysms was evaluated with receiver operating characteristic (ROC) methodology. The negative effect of JPEG compression on automated detection was observed even at levels of compression sometimes used in retinal eye-screening programmes and these may have important clinical implications for deciding on acceptable levels of compression for a fully automated eye-screening programme.

Tracking involves estimating not only the global motion but also local perturbations or deformations corresponding to a specified object of interest. From this, motion can be decoupled into a finite dimensional state space (the global motion) and the more interesting infinite dimensional state space (deformations). Recently, the incorporation of the particle filter with geometric active contours which use first and second moments has shown robust tracking results. By generalizing the statistical inference to entire probability distributions, we introduce a new distribution metric for tracking that is naturally able to better model the target. Also, due to the multiple hypothesis nature of particle filtering, it can be readily seen that if the background resembles the foreground, then one might lose track. Even though this can be described as a finite dimensional problem where global motion can be modeled and learned online through a filtering process, we approach this task by incorporating a separate energy term in the deformable model that penalizes large centroid displacements. Robust results are obtained and demonstrated on several surveillance sequences.

This paper addresses methods for statistical uncertainty analysis to determine the measurement accuracy associated with a video-extensometer system. Two different approaches for statistical uncertainty analysis - a purely statistical and an analytical approximation - are presented. The statistical method is based on evaluation of images acquired at conditions of repeatability; whereas the analytical approach consists of application of the law of first order error propagation to the particular processing steps of the evaluation procedure. The derivation of the law of first order error propagation is briefly revised in order to emphasize possible sources of error caused by its application. Moreover, the computation of the Jacobian matrix required for first order approximations of error propagation is illustrated for explicit and implicit vector-valued functions as well as for linear least squares problems as this represents a task typically arising in metric vision applications. Finally, the two approaches are applied to the specific processing steps for the evaluation of the images acquired with the video-extensometer system. Comparison of the results obtained with the different methods show negligible deviations, proving the application of the law of first order error propagation to be a suitable means to analytically estimate statistical uncertainty.

Augmented reality is used to improve color segmentation on human body or on precious no touch artifacts. We propose a technique to project a synthesized texture on real object without contact. Our technique can be used in medical or archaeological application. By projecting a suitable set of light patterns onto the surface of a 3D real object and by capturing images with a camera, a large number of correspondences can be found and the 3D points can be reconstructed. We aim to determine these points of correspondence between cameras and projector from a scene without explicit points and normals. We then project an adjusted texture onto the real object surface. We propose a global and automatic method to virtually texture a 3D real object.

Infrared imaging is based on the measurement of the radiation of an object and its conversion to temperature. A very important parameter of the conversion procedure is emissivity, which defines the capability of a material to absorb and radiate energy. For most applications, emissivity is assumed to be constant. However, when infrared images are taken from steel strips in an industrial environment, where the measurement is influenced by thermal reflections of surrounding objects, the consideration of emissivity as a constant can lead to large errors in temperature measurement. To overcome problems generated by variations in emissivity, one solution is to measure temperature where the steel strip forms a wedge, acting as a cavity. In the deepest part of the wedge, emissivity is 1, making the emissivity problems disappear. This work presents a real-time image processing system to acquire infrared line scans for steel strips using the wedge method. The proposed system deals with two main problems: infrared line scans must be extracted in real-time from the deepest part of the wedge in the rectangular infrared images, and pixels belonging to the line scan must be translated to real-world units.

This paper proposes a new vision-based fire detection method for real-life application. Most previous vision-based methods using color information and temporal variations of pixels produce frequent false alarms due to the use of many heuristic features. Plus, there is usually a computation delay for accurate fire detection. Thus, to overcome these problems, candidate fire regions are first detected using a background model and color model of fire. Probabilistic models of fire are then generated based on the fact that fire pixel values in consecutive frames change constantly and these models are applied to a Bayesian Network. This paper uses a three-level Bayesian Network that contains intermediate nodes, and uses four probability density functions for evidence at each node. The probability density functions for each node are modeled using the skewness of the color red and three high frequency components obtained from a wavelet transform. The proposed system was successfully applied to various fire-detection tasks in real-world environments and effectively distinguished fire from fire-colored moving objects.

Injection molded plastic parts are often influenced with the surface defect tiger stripes, which dramatically reduce the visual quality. Tiger stripes are known as alternating bands of bright and dull regions normally to the molded flow direction. This defect highly depends on the injection time and on the formation of the plastic compound. In the last years, the intensity of the tiger stripes defect was controlled visually. For quantifying the tiger strip defect a new, efficient, repeatable, reliable and nondestructive optical measurement system is proposed. To evaluate the dependency of the injection time, a number of five DIN-A5 plastic specimens are molded. Each of the five plates consists of the same material but they have different injection times. For the measurement, one specimen is put into the specimen holder, which is placed on the drawer of a closed cabinet. In this inside black painted cabinet a LED light source and a CCD Camera are mounted. The beams of the LED light are diffuse reflected on the surface of the specimen. To catch only parallel beams by the lens of the camera a large distance between specimen and camera is realized by two justified mirrors in the cabinet. The bright and dull regions of the tiger stripe defect have different diffuse reflection parameters. Thus in a picture of defined brightness the visibility of this defect is very good. To enhance the repeatability the failure of the camera noise and of the light oscillation is reduced by mends of averaging multiple images. Next, the surface structure is filtered out of the image and a representing number of horizontal grey-value lines are extracted. The so called tiger line signal is the difference between the grey line and a calculated polynomial function (degree of 6) and shows the surface defect of each line oscillating on the zero x-axis. For each tiger line signal the mean squared error is evaluated. To calculate a quantitative value of the whole surface, all line errors are averaged to the so called MSE-value. Measurements and comparisons show, that this MSE-value represents surface defects and especially the intensity of tiger stripes very good. The repeating error is lower than 1 %. Experiments for showing unknown effects of normal and of accelerated aging and weathering of plastic surfaces were done successfully.

The thin film transistor liquid crystal display (TFT-LCD) has become an actively used front of panel display technology with an increasing market. Intrinsically there is a region of non uniformity with low contrast that to human eye is perceived as a defect. Because the grey-level difference between the defect and the background is small, the conventional edge detection techniques are hardly applicable to detect these low contrast defects. Although several effort were dedicated in classifying the patterned TFT-LCD defects, only few researches were conducted on detecting the unpatterned TFT-LCD defects that accounts for approximately 15% of all defects produced during the manufacturing stages. This paper proposes a detection method for the un-patterned TFT-LCD defects by using the directional filter bank (DFB), Shen-Castan filter and maximum Feret's diameter. The effectiveness of the proposed method is tested through the experiment using real TFT-LCD panel images.

Detecting defects is important technology of the TFT-LCD (Thin Film Transistor-Liquid Crystal Display) production process for quality control. For high quality and improving productive, defect detection is performed on each manufacturing process. In array process, defect inspection is divided into inspection for active matrix area and inspection for pad area. Inspection on active matrix area has used period of pattern to detect defect. As pad has non repetitive pattern, period can not be used for defect detection. Therefore, defects on pad have been detected by referential method comparing to pre-stored reference pad image. Subtraction has been used for comparison with reference pad. This method is problematic for pad defect inspection due to variance in the shapes of pad, illumination change and alignment error. In this paper, we propose the inspection method making up for limitation of referential method which has been used for TFT-LCD pad. Inspection is performed by applying morphological method to each horizontal line. By finding valley of each line, defect is detected.

For activities of agronomical research institute, the land experimentations are essential and provide relevant information on crops such as disease rate, yield components, weed rate... Generally accurate, they are manually done and present numerous drawbacks, such as penibility, notably for wheat ear counting. In this case, the use of color and/or texture image processing to estimate the number of ears per square metre can be an improvement. Then, different image segmentation techniques based on feature extraction have been tested using textural information with first and higher order statistical methods. The Run Length method gives the best results closed to manual countings with an average error of 3%. Nevertheless, a fine justification of hypothesis made on the values of the classification and description parameters is necessary, especially for the number of classes and the size of analysis windows, through the estimation of a cluster validity index. The first results show that the mean number of classes in wheat image is of 11, which proves that our choice of 3 is not well adapted. To complete these results, we are currently analysing each of the class previously extracted to gather together all the classes characterizing the ears.

In this paper two in-line measurement systems for the geometric inspection of welded or profiled steel strips are presented. For the first case, the welded strips, the paper presents a light sectioning based measurement system for the in-line inspection of the welding seam. The system measures the vertical and angular misalignment of the welded parts in-line, right after the welding station. Based on the calculated offset of the two welded parts on one hand the welding adjustment can be optimized. On the other hand, quality relevant data is collected. The inspection of the vertical offset is done with an accuracy of ±5μm (±2σ). In addition to the measurement of the vertical and angular offset, the system also inspects the width of the welding seam and the seam's shape and depth, all three with an accuracy of ±15μm (±2σ). The second system presented in the paper is a measurement system for the inspection of the cross section of profiled strips. The system consists of three light sectioning based measurement heads and is mounted right after the profiling plates. Regarding the measurement precision, the shape of the profiled strip can be measured with an accuracy of better than ±10μm (±2σ). The paper for both applications describe the sensor setup, geometric conditions, the used hardware and the calibration and registration approaches.

The human object segmentation and classification are main work in the applications of Intelligent Visual Surveillance System or Passenger Flow Counting System. Traditional approaches to segment and classify human objects are usually based on the face, leg motion and silhouette. These algorithms' performances and their applications have proved to be effective in recent years. But these algorithms all assume that features can always be extracted. In complex situations, however, features adopted in traditional algorithms might not be extracted, because human attitude and illumination change greatly. In this case, if a definite feature is used, the algorithm's accuracy will fall. In this paper we propose an approach to select the feature and the corresponding algorithm adaptively based on the human attitude and object neighborhood illumination. The selected features can be used in the following tracking operation. Because this method solves the human object segmentation and classification problem, it can broad the 3D recovery and behavior understanding research results in simple situations to the application in complex situations. In this paper, the algorithms are proposed for the human attitude and illumination detection, the feature selection strategies in different situation are given. The experimental results show that the algorithm can detect the object lightness properly, and can give the right attitude for feature selection. The algorithms have good performance and computation efficiency.

In this paper, we propose a method for detecting unusual human behavior using monocular camera which is not moving. Our system composed of three modules which are moving object detection, tracking, and event recognition. The key part is event recognition module. We define unusual events which are composed of two simple events (drop off luggage, unattended luggage) and two complex events (abandoned luggage and steal luggage). In order to detect the simple event, we construct Bayesian network in each unusual event. We extract evidences using bounding box properties which are the location of moving objects, speed, distance between the person and the other moving object (such as bag), existing time. And then, we use finite state automaton which shows the temporal relation of two simple events to detect complex events. To evaluate the performance, we compare the frame number when an even is triggered with our results and the ground truth. The proposed algorithm showed good results on the real world environment and also worked at real time speed.

In this paper, we present a hardware implementation of a face detector for surveillance applications. To come up with a computationally cheap and fast algorithm with minimal memory requirement, motion and skin color information are fused successfully. More specifically, a newly appeared object is extracted first by comparing average Hue and Saturation values of background image and a current image. Then, the result of skin color filtering of the current image is combined with the result of a newly appeared object. Finally, labeling is performed to locate a true face region. The proposed system is implemented on Altera Cyclone2 using Quartus II 6.1 and ModelSim 6.1. For hardware description language (HDL), Verilog-HDL is used.

This paper proposes an effective approach for online inspecting and recognizing the assembly structure inside three-dimensional objects using multiple views technique and X-ray digital radiography system. During the offline study process, the paper obtains a gray image sequence of a standard sample in multiple circumferential orientations. Utilizing the idea of classifying identification, the paper locates and extracts different characters of different parts in each image of the sequence and establishes corresponding character sequence libraries. In online detection stage, the program finds the optimum solutions to all different target parts in the library with bisearch method and carries out exactness image matching with correlation coefficient weighted of multi-character via Bayes decision. Aiming at the issue of some objects may be occluded by others in a scene, the paper puts forward to rotate the product some certain angles and re-match. Furthermore, the paper analyzes the relationships of misjudgment ratio with product assembling tolerance, the size of target part and identifying velocity. Based on this approach, the first domestic X-ray automatism detection system has been developed and it is successfully applied in online detecting some axis symmetric products which assembly structures inside are complex.

In this paper, we present an algorithm of estimating new-view vehicle speed. Different from far-view scenario, near-view image provides more specific vehicle information such as body texture and vehicle identifier which makes it practical for individual vehicle speed estimation. The algorithm adopts the idea of Vanishing Point to calibrate camera parameters and Gaussian Mixture Model (GMM) to detect moving vehicles. After calibrating, it transforms image coordinates to the real-world coordinates using a simple model - the Pinhole Model and calculates the vehicle speed in real-world coordinates. Adopting the idea of Vanishing Point, this algorithm only needs two pre-measured parameters: camera height and distance between camera and middle road line, other information such as camera orientation, focal length, and vehicle speed can be extracted from video data.