<p>En modern grafikprocessor är oerhört kraftfull och har en prestanda som potentiellt sett är många gånger högre än för en modern mikroprocessor. I takt med att grafikprocessorn blivit alltmer programmerbar har det blivit möjligt att använda den för beräkningstunga tillämpningar utanför dess normala användningsområde. Inom det här arbetet utreds vilka möjligheter och begränsningar som uppstår vid användandet av grafikprocessorer för generell programmering. Arbetet inriktas främst mot signal- och bildbehandlingstillämpningar men mycket av principerna är tillämpliga även inom andra områden.</p><p>Ett ramverk för bildbehandling implementeras och några algoritmer inom bildanalys realiseras och utvärderas, bland annat stereoseende och beräkning av optiskt flöde. Resultaten visar på att vissa tillämpningar kan uppvisa en avsevärd prestandaökning i en grafikprocessor jämfört med i en mikroprocessor men att andra tillämpningar kan vara ineffektiva eller mycket svåra att implementera.</p> / <p>The modern graphical processing unit, GPU, is an extremely powerful unit, potentially many times more powerful than a modern microprocessor. Due to its increasing programmability it has recently become possible to use it in computation intensive applications outside its normal usage. This work investigates the possibilities and limitations of general purpose programming on GPUs. The work mainly concentrates on signal and image processing although much of the principles are applicable to other areas as well.</p><p>A framework for image processing on GPUs is implemented and a few computer vision algorithms are implemented and evaluated, among them stereo vision and optical flow. The results show that some applications can gain a substantial speedup when implemented correctly in the GPU but others can be inefficent or extremly hard to implement.</p>

Image analysis problems, posed mathematically as variational principles or as partial differential equations, are amenable to numerical solution by relaxation algorithms that are local, iterative, and often parallel. Although they are well suited structurally for implementation on massively parallel, locally-interconnected computational architectures, such distributed algorithms are seriously handicapped by an inherent inefficiency at propagating constraints between widely separated processing elements. Hence, they converge extremely slowly when confronted by the large representations necessary for low-level vision. Application of multigrid methods can overcome this drawback, as we established in previous work on 3-D surface reconstruction. In this paper, we develop efficient multiresolution iterative algorithms for computing lightness, shape-from-shading, and optical flow, and we evaluate the performance of these algorithms on Synthetic images. The multigrid methodology that we describe is broadly applicable in low-level vision. Notably, it is an appealing strategy to use in conjunction with regularization analysis for the efficient solution of a wide range of ill-posed visual reconstruction problems.

In the task of human action recognition in uncontrolled video, motion features are used widely in order to achieve subject and appearence invariance. We implemented 3 Histograms of Oriented Optical Flow based method which have a common motion feature extraction phase. We compute an optical flow field over each frame of the video. Then those flow vectors are histogrammed due to angle values to represent each frame with a histogram. In order to capture local motions, The bounding box of the subject is divided into grids and the angle histograms of all grids are concetanated to obtain the final motion feature vector. Motion Features are supplied to 3 dierent classification system alternatives containing clustering combined with HMM, clustering with K-nearest neighbours and average histograms methods. Three methods are implemented and results are evaluated over Weizmann and KTH datasets.

It has been well known that the architecture of insect compound eyes contributes outstanding capability for precise and efficient observation of moving objects. If this technique can be transferred to the domain of engineering applications, significant improvement on visual tracking of moving objects will be greatly expected. The brightness variation, caused by relative velocity of the camera and environment in a sequence of images, is called optical flow. The advantage of the optical-flow-based visual servo methods is that features of the moving object do not have to be known in advance. Therefore, they can be applied for general positioning and tracking tasks. The purpose of this thesis is to develop a visual servo system with trinocular cameras. For mimicking the configuration of compound eyes of insects, the arrangement of the divergent trinocular cameras is applied. In order to overcome possible difficulties of unknown or uncertain parameters, an image servo technique using the robust discrete-time sliding-mode control algorithm to track an object moving in 2D space is developed.

The car has become bigger, faster and more advanced for each passing year since its first appearance, and the safety requirements have also become stricter. Computer vision based support is a growing area of safety features where the car is equipped with a mono- or stereo camera. It can be used for detecting pedestrians walking out in the street, give a warning for wild-life during a cold January night using night-vision cameras and much more. This master thesis investigates the problem of detecting and tracking overtaking vehicles. Vehicles that overtake are only partly visible in the beginning, rendering it hard for standard detection/classification algorithms to get a positive detection. The need to quickly detect an incoming vehicle is crucial to be able to take fast counter-measure, such as braking, if needed. A novel approach referred to as the \textit{Wall detector} is suggested, detecting incoming vehicles using one-dimensional optical flow. Under the assumption that an overtaking car is moving in parallel to the ego-vehicle, both cars are moving towards the vanishing point in the image. A detection wall, consisting of several detection lines moving towards the vanishing point, is created, making all objects that are moving parallel to the ego-vehicle move along these lines. The result is a light-weight and fast detector with good detection performance in real-time. Several approaches for the Wall detector are implemented and evaluated, revealing that a feature based approach is the best choice. The information from the system can be used as input to heavier algorithms, boosting the confidence or to initialize a track.

This dissertation presents a technique for recovering translational motion parameters using two simplified planar compound-like eye schemes, namely a parallel trinocular system and a single-row Superposition-type Planar Compound-like Eye (SPCE). In the parallel trinocular scheme, a least squares estimation algorithm is developed for recovering the translational motion parameters. The proposed approach resolves the matrix singularity problem encountered when attempting to recover motion parameters using a conventional binocular scheme. To further reduce the computational complexity of the motion estimation process, a compact closed-form scheme is also proposed to estimate the translational motion parameters. The closed-form algorithm not only resolves the matrix singularity problem, but also avoids the requirement for matrix manipulation. As a result, it has a low computational complexity and is therefore an ideal solution for performing motion estimation in complex, real-world visual imaging applications following an initial image filtering process. The performance of the closed-form algorithm is evaluated by performing a series of numerical simulations in which translational displacements of various magnitudes in three-dimensional space are recovered in both noise-free and perturbed environments. In general, the results demonstrate that the translational motion parameters can be reconstructed with a high degree of accuracy provided that the motion in the depth direction is limited to small displacements only. Having developed a motion estimation scheme for a parallel trinocular system, additional charge coupled device (CCD) cameras are added in the horizontal direction to create a single-row SPCE. Translational motion models for the SPCE are then constructed by stacking the optical flow equations in the horizontal direction. The ego-translational parameters are then extracted using a simple least squares estimation algorithm. The simulation results reveal that the introduction of additional cameras to the machine vision system ensures an excellent motion estimation performance without the need for filters of any kind even when the viewing field is characterized by significant noise or the CCD deployment within the SPCE configuration has a non-uniform distribution. Overall, the parallel binocular scheme and single-row SPCE configuration presented in this dissertation demonstrate a high degree of robustness toward noise and enable the motion estimation process to be performed in a rapid and computationally efficient manner using a simple least squares approximation approach. Whilst science can not realistically hope to improve upon the visioning capabilities found in the insect world, the techniques presented in this dissertation nonetheless provide a sound foundation for the development of artificial planar-array compound-like eyes which mimic the mechanisms at work in biological compound eyes and attain an enhanced visioning performance as a result.

As a result of everlasting developments in multimedia technologies, all kinds of objects tracking theory using machine vision or image process methods have been proposed. Most of the methods are based on shape of the object. For this reason, the profile of the tracked object must be known in advance. In many situations, we expect to track the object whose shape is unknown but speed or direction is explicit. For instance, speed or moving direction of the object is known. This thesis presents a spatio-temporal tracking technique, which extracts image information depending on speed of the moving object regardless of its shape. Furthermore, combination of the proposed method in spatio-temporal domain and the optical flow scheme makes the whole tracking system even more robust.

This research seeks to expand the role of variational and adjoint processing methods into segments of the sonar, radar, and nondestructive testing communities where they have not yet been widely introduced. First, synthetic aperture reconstruction is expressed in terms of the adjoint operator. Many, if not all, practical imaging modalities can be traced back to this general result, as the adjoint is the foundation for backprojection-type algorithms. Next, active surfaces are developed in the context of the Helmholtz equation for the cases of opaque scatterers (i.e., with no interior field) embedded in free space, and penetrable scatterers embedded in a volume which may be bounded. The latter are demonstrated numerically using closed-form solutions based on spherical harmonics. The former case was chosen as the basis for a laboratory experiment using Lamb waves in an aluminum plate. Lamb wave propagation in plates is accurately described by the Helmholtz equation, where the field quantity is the displacement potential. However, the boundary conditions associated with the displacement potential formulation of Lamb waves are incompatible with the shape gradient derived for the Helmholtz equation, except for very long or very short wavelengths. Lastly, optical flow is used to solve a new and unique problem in the field of synthetic aperture sonar. Areas of acoustic focusing and dilution attributable to refraction can sometimes resemble the natural bathymetry of the ocean floor. The difference is often visually indistinguishable, so it is desirable to have a means of detecting these transient refractive effects without having to repeat the survey. Optical flow proved to be effective for this purpose, and it is shown that the parameters used to control the algorithm can be linked to known properties of the data collection and scattering physics.

Research on video-based face recognition has started getting increased attention in the past few years. Algorithms developed for video have an advantage from the availability of plentitude of frames in videos to extract information from. Despite this fact, most research in this direction has limited the scope of the problem to the application of still image-based approaches to some selected frames on which 2D algorithms are expected to perform well. It can be realized that such an approach only uses the spatial information contained in video and does not incorporate the temporal structure.Only recently has the intelligence community begun to approach the problem in this direction. Video-based face recognition algorithms in the last couple of years attempt to simultaneously use the spatial and temporal information for the recognition of moving faces. A new face recognition method that falls into the category of algorithms that adopt spatio-temporal representation and utilizes dynamic information extracted from video is presented. The method was designed based on the hypothesis that the strain pattern exhibited during facial expression provides a unique "fingerprint" for recognition. First, a dense motion field is obtained with an optical flow algorithm. A strain pattern is then derived from the motion field. In experiments with 30 subjects, results indicate that strain pattern is an useful biometric, especially when dealing with extreme conditions such as shadow light and face camouflage, for which conventional face recognition methods are expected to fail. The ability to characterize the face using the elastic properties of facial skin opens up newer avenues to the face recognition community in the context of modeling a face using features beyond visible cues.

Considering todays growing society and developing technologies which are co-influential between each other, there is a larger scope of security concerns, traffic congestion due to improper planning and hence a greater need of more intelligent video surveillance. In this thesis, we have worked on developing such intelligent video surveillance system which mainly focusses on cell area such as parking spaces. The system operates on outdoor environment with a stationary camera; the main objective of this system is detecting and tracking of moving objects mainly cars. Two detection algorithms were developed using optical flow as core strategy. In the first algorithm the flow vectors were classified based on their magnitude and orientation; the GOMAG algorithm. The second algorithm used K-means method on the flow vectors to achieve the classification for moving object detection; the SKMO algorithm. A comparison analysis was done between the proposed algorithms and well known detection algorithms of background modeling and Otsu’s segmentation of flow vectors. The both proposed algorithms performed significantly better than background modeling and Otsu’s segmentation of flow vectors algorithms. The SKMO algorithm showed better stability and processed time efficiency than the GOMAG algorithm.