Burak Benligiray

Video-Based Lane Detection Using a Fast Vanishing Point Estimation Method

Lane detection algorithms constitute a basis for intelligent vehicle systems such as lane tracking and involuntary lane departure detection. In this paper, we propose a simple and video-based lane detection algorithm that uses a fast vanishing point estimation method. The first step of the algorithm is to extract and validate the line segments from the image with a recently proposed line detection algorithm. In the next step, an angle based elimination of line segments is done according to the perspective characteristics of lane markings. This basic operation removes many line segments that belong to irrelevant details on the scene and greatly reduces the number of features to be processed afterwards. Remaining line segments are extrapolated and superimposed to detect the image location where majority of the linear edge features converge. The location found by this efficient operation is assumed to be the vanishing point. Subsequently, an orientation-based removal is done by eliminating the line segments whose extensions do not intersect the vanishing point. The final step is clustering the remaining line segments such that each cluster represents a lane marking or a boundary of the road (i.e. sidewalks, barriers or shoulders). The properties of the line segments that constitute the clusters are fused to represent each cluster with a single line. The nearest two clusters to the vehicle are chosen as the lines that bound the lane that is being driven on. The proposed algorithm works in an average of 12 milliseconds for each frame with 640×480 resolution on a 2.20 GHz Intel CPU. This performance metric shows that the algorithm can be deployed on minimal hardware and still provide real-time performance.

A robust CSS corner detector based on the turning angle curvature of image gradients

In this study, we present a new contour-based corner detection method based on the turning angle curvature computed from the contour gradients of the image. In general, curvature is computed with the pixel locations of the extracted image contours. In most contour extraction methods, the image gradient information is already computed. The proposed algorithm makes use of this available information to compute the curvature function and takes local extremums as potential corner candidates. Afterwards, the candidates are validated by a novel validation algorithm which tries to approximate the local geometric structure of the contour with an iterative least squares estimation algorithm. Thus, we not only eliminate the false detected corners; but also estimate the corner strength precisely in terms of degrees. The experiments show that the detected corners with gradient-based turning angle curvature are more durable to affine transformations according to the ACU and LE criterions.

Combining feature-based and model-based approaches for robust ellipse detection

Fast and robust ellipse detection is a vital step in many image processing and computer vision applications. Two main approaches exist for ellipse detection, i.e., model-based and feature-based. Model-based methods require much more computation, but they can perform better in occlusions. Feature-based approaches are fast but may perform insufficient in cluttered cases. In this study, we propose an hybrid method which combines both approaches to accelerate the process without compromising accuracy. We extract elliptical arcs to narrow down search space by obtaining seeds for prospective ellipses. For each seed arc, we compute a limited search region consisting of hypothetical ellipses that each can be formed with that seed. Later, we vote them on the edge image to determine best hypothesis among the all, if exists. We tested the proposed algorithm on a public dataset and promising results are obtained compare to state of the art methods in the literature.

On the efficiency issues of virtual keyboard design

Virtual keyboards are useful tools, which ease effortless entry of textual data and enable typing with alternative input hardware such as a single switch. Early virtual keyboards are designed similar to physical keyboards in terms of appearance. Since a physical keyboard is designed for tactile-use; the usability of the first virtual keyboards comes up short in utilization with pointing devices. For this reason, more useful virtual keyboards are proposed with improvements in modal and functional properties. In this study, we examine a couple of design issues for virtual keyboards to provide their efficient utilization with pointing devices. We analyze the effects of visual key layout to the performance of virtual keyboards in connection with the statistical properties of the target languages vocabulary. We also propose a virtual keyboard design for a comfortable text entry experience based on our observations.

Sequential forward feature selection for facial expression recognition

Facial expression recognition is an important computer vision problem with various applications. In this study, we investigate the effectiveness of features derived from facial landmarks in facial expression recognition. Distances between two combinations of facial landmarks constitute a distance vector. Features we use are the changes in the distance vectors extracted from expressive and neutral states of the face. The obtained feature vector contains elements that are relatively useless in expression recognition. By applying forward sequential feature selection, a subset of the most effective elements is formed. The chosen features are classified using a multi-class support vector machine. The performance of the proposed method is measured using Extended Cohn-Kanade dataset with seven expressions (anger, contempt, disgust, fear, happy, sad and surprised) and resulted in 89.9% mean class recognition accuracy.

Lens Distortion Rectification Using Triangulation Based Interpolation

Nonlinear lens distortion rectification is a common first step in image processing applications where the assumption of a linear camera model is essential. For rectifying the lens distortion, forward distortion model needs to be known. However, many self-calibration methods estimate the inverse distortion model. In the literature, the inverse of the estimated model is approximated for image rectification, which introduces additional error to the system. We propose a novel distortion rectification method that uses the inverse distortion model directly. The method starts by mapping the distorted pixels to the rectified image using the inverse distortion model. The resulting set of points with subpixel locations are triangulated. The pixel values of the rectified image are linearly interpolated based on this triangulation. The method is applicable to all camera calibration methods that estimate the inverse distortion model and performs well across a large range of parameters.

Greedy search for descriptive spatial face features

Facial expression recognition methods use a combination of geometric and appearance-based features. Spatial features are derived from displacements of facial landmarks, and carry geometric information. These features are either selected based on prior knowledge, or dimension-reduced from a large pool. In this study, we produce a large number of potential spatial features using two combinations of facial landmarks. Among these, we search for a descriptive subset of features using sequential forward selection. The chosen feature subset is used to classify facial expressions in the extended Cohn-Kanade dataset (CK+), and delivered 88.7% recognition accuracy without using any appearance-based features.

A comparison of classification methods for local binary patterns

Texture recognition is an important tool used for content-based image retrieval, face recognition, and satellite image classification applications. One of the most successful features for texture recognition is local binary patterns (LBP), which computes local intensity differences for a pixel with respect to its neighbor pixels. In many studies in the literature, histogram based similarity measures are employed to classify LBP features. In this study, we investigate the performance of support vector machines, linear discriminant analysis, and linear regression classifier to improve the success of LBP features. We achieved 84.4% classification success using linear regression classification.

HEp-2 cell classification using a deep neural network trained for natural image classification

Deep convolutional neural networks is a recently developed method that yields very successful results in image classification. Deep neural networks, which have a high number of parameters, require a large amount of data to avoid overfitting during training. For applications in which the available data is not adequate to train a deep neural network from the scratch, deep neural networks trained for similar objectives can be used as a starting point. In this study, cell images are classified using a deep neural network trained to classify objects in natural images. Even though classification of natural images and cell images are very different objectives, cell images are able to be classified with 74.1% mean class accuracy. The results show that features used for visual classification by deep convolutional neural networks may be more universal than assumed.

Blind rectification of radial distortion by line straightness

Lens distortion self-calibration estimates the distortion model using arbitrary images captured by a camera. The estimated model is then used to rectify images taken with the same camera. These methods generally use the fact that built environments are line dominated and these lines correspond to lines on the image when distortion is not present. The proposed method starts by detecting groups of lines whose real world correspondences are likely to be collinear. These line groups are rectified, then a novel error function is calculated to estimate the amount of remaining distortion. These steps are repeated iteratively until suitable distortion parameters are found. A feature selection method is used to eliminate the line groups that are not collinear in the real world. The method is demonstrated to successfully rectify real images of cluttered scenes in a fully automatic manner.