As research into autonomous vehicles and Advanced Driver Assistance Systems (ADAS) has grown, research into computer vision techniques to detect objects and lane lines within images has also grown. The heavier computational load of modern techniques involving neural net- works and machine learning limits the ability to downscale to cheaper, less computationally- capable platforms when needed. The goal of the project is to develop a robust and computationally efficient method to estimate vehicle position within a lane. A clothoid lane line model based in real-world coor- dinates is projected into the image pixel-space where a novel approach to image segmentation and active contour modeling is performed. Another novel approach presented is the use of velocity as an input from a source outside the algorithm into the process to predict the initial conditions of the model in the next frame, rather than using the algorithm to produce an estimate of the velocity as an output to other systems. Validation is performed using the TuSimple dataset using both ideal and realistic scenarios to evaluate the performance of the various aspects of the algorithm against the current state-of-the-art methods. / Master of Science / As interest grows in autonomous driving and systems used to assist drivers on the road, many techniques have been developed to identify objects of interest in the surrounding envi- ronment. One of the most common involves neural networks that use images from cameras to identify targets such as pedestrians, signs, and lane lines. Lane lines are particularly of interest as vehicle control systems need information on where the car is on the road in order to properly stay in the lanes. One major downside of the current implementation of neural networks is that they require more powerful computers and often cannot run on many cheaper, less capable machines. This study proposes a method that focuses on using small sections of the image rather than the entire picture in order to run the algorithm in a computationally efficient manner. A road model is used to keep track of where the lane is in the image and is updated as new camera images are provided. This method presents a novel way of selecting information in the image and using the speed of the car to estimate where the car will be in the next camera frame. To evaluate the performance of the algorithm, a public dataset of videos is used to run the algorithm and compare against other methods that have used the same set of videos.

Computational vision and image analysis is a multidisciplinary scientific field that aims to make computers "see" in a way that is comparable to human perception. It is currently one of the most challenging research areas in artificial intelligence. In this regard, the extraction of information from the vast amount of visual data that are available today as well as the exploitation of the resulting information space becomes one of the greatest challenges in our days. To address such a challenge, this thesis describes a very general computational framework that can be used for performing efficient inference and learning for visual perception based on very rich and powerful models.

Computer vision approximates human vision using computers. Two subsets are explored in this work: image segmentation and visual tracking. Segmentation involves partitioning an image into logical parts, and tracking analyzes objects as they change over time. The presented research explores a key hypothesis: localizing analysis of visual information can improve the accuracy of segmentation and tracking results. Accordingly, a new class of segmentation techniques based on localized analysis is developed and explored. Next, these techniques are applied to two challenging problems: neuron bundle segmentation in diffusion tensor imagery (DTI) and plaque detection in computed tomography angiography (CTA) imagery. Experiments demonstrate that local analysis is well suited for these medical imaging tasks. Finally, a visual tracking algorithm is shown that uses temporal localization to track objects that change drastically over time.

This thesis aims to contribute to the area of visual tracking, which is the process of identifying an object of interest through a sequence of successive images. The thesis explores kernel-based statistical methods, which map the data to a higher dimensional space. A pre-image framework is provided to find the mapping from the embedding space to the input space for several manifold learning and dimensional learning algorithms. Two algorithms are developed for visual tracking that are robust to noise and occlusions. In the first algorithm, a kernel PCA-based eigenspace representation is used. The de-noising and clustering capabilities of the kernel PCA procedure lead to a robust algorithm. This framework is extended to incorporate the background information in an energy based formulation, which is minimized using graph cut and to track multiple objects using a single learned model. In the second method, a robust density comparison framework is developed that is applied to visual tracking, where an object is tracked by minimizing the distance between a model distribution and given candidate distributions. The superior performance of kernel-based algorithms comes at a price of increased storage and computational requirements. A novel method is developed that takes advantage of the universal approximation capabilities of generalized radial basis function neural networks to reduce the computational and storage requirements for kernel-based methods.

Due to the large extent to which important biological mechanisms are conserved evolutionarily, the study of a simple soil nematode, C. elegans, has provided the template for significant advances in biology. Use of this model organism has accelerated in recent years as developments of advanced reagents such as synapse localized fluorescent markers have provided powerful tools to study the complex process of synapse formation and remodeling. Even as much routine biology work, such as sequencing, has become faster and easier, imaging protocols have remained essentially unchanged over the past forty years of research. This, coupled with the ability to visualize small, complex features as a result of new fluorescent reagents, has resulted in genetic screens in C. elegans becoming increasingly labor intensive and slow because microscopy mainly relies on manual mounting of animals and phenotyping is usually visually done by experts. Genetic screens have become the rate limiting factor for much of modern C. elegans research. Furthermore, phenotyping of fluorescent expression has remained a primarily qualitative process which has prevented statistical analysis of subtle features. To address these issues, a comprehensive system to allow autonomous screening for novel mutants was created. This was done by developing novel microfluidic devices to enable high-throughput screening, systems-level components to allow automated operation, and a computer vision framework for identification and quantitative phenotyping of synaptic patterns. The microfluidic platform allows for imaging and sorting of thousands of animals at high-magnification within hours. The computer vision framework employs a two-stage feature extraction to incorporate local and regional features and allows for synapse identification in near real-time with an extremely low error rate. Using this system thousands of mutagenized animals were screened to indentify numerous novel mutants expressing altered synaptic placement and development. Fully automated screening and analysis of subtle fluorescent phenotypes will allow large scale RNAi and drug screens. Combining microfluidics and computer vision approaches will have a significant impact on the biological community by removing a significant bottleneck and allowing large-scale screens that would have previously been too labor intensive to attempt.

As demand for real-time image processing increases, the need to improve the efficiency of image processing systems is growing. The process of image segmentation is often used in preprocessing stages of computer vision systems to reduce image data and increase processing efficiency. This dissertation introduces a novel image segmentation approach known as leap segmentation, which applies a flexible definition of adjacency to allow groupings of pixels into segments which need not be spatially contiguous and thus can more accurately correspond to large surfaces in the scene. Experiments show that leap segmentation correctly preserves an average of 20% more original scene pixels than traditional approaches, while using the same number of segments, and significantly improves execution performance (executing 10x - 15x faster than leading approaches). Further, leap segmentation is shown to improve the efficiency of a high-level vision application for scene layout analysis within 3D scene reconstruction. The benefits of applying image segmentation in preprocessing are not limited to single-frame image processing. Segmentation is also often applied in the preprocessing stages of video analysis applications. In the second contribution of this dissertation, the fast, single-frame leap segmentation approach is extended into the temporal domain to develop a highly-efficient method for multiple-frame segmentation, called video leap segmentation. This approach is evaluated for use on mobile platforms where processing speed is critical using moving-camera traffic sequences captured on busy, multi-lane highways. Video leap segmentation accurately tracks segments across temporal bounds, maintaining temporal coherence between the input sequence frames. It is shown that video leap segmentation can be applied with high accuracy to the task of salient segment transformation detection for alerting drivers to important scene changes that may affect future steering decisions. Finally, while research efforts in the field of image segmentation have often recognized the need for efficient implementations for real-time processing, many of today’s leading image segmentation approaches exhibit processing times which exceed their camera frame periods, making them infeasible for use in real-time applications. The third research contribution of this dissertation focuses on developing fast implementations of the single-frame leap segmentation approach for use on both single-core and multi-core platforms as well as on both high-performance and resource-constrained systems. While the design of leap segmentation lends itself to efficient implementations, the efficiency achieved by this algorithm, as in any algorithm, is can be improved with careful implementation optimizations. The leap segmentation approach is analyzed in detail and highly optimized implementations of the approach are presented with in-depth studies, ranging from storage considerations to realizing parallel processing potential. The final implementations of leap segmentation for both serial and parallel platforms are shown to achieve real-time frame rates even when processing very high resolution input images. Leap segmentation’s accuracy and speed make it a highly competitive alternative to today’s leading segmentation approaches for modern, real-time computer vision systems.

Thesis (Ph.D.). / Written for the Dept. of Electrical & Computer Engineering. Title from title page of PDF (viewed 2008/01/12). Includes bibliographical references.

Gaze tracking is the estimation of the point in space a person is “looking at”. This is widely used in both diagnostic and interactive applications, such as visual attention studies and human-computer interaction. The most common commercial solution used to track gaze today uses a combination of infrared illumination and one or more cameras. These commercial solutions are reliable and accurate, but often expensive. The aim of this thesis is to construct a simple single-camera gaze tracker from off-the-shelf components. The method used for gaze tracking is based on infrared illumination and a schematic model of the human eye. Based on images of reflections of specific light sources in the surfaces of the eye the user’s gaze point will be estimated. Evaluation is also performed on both the software and hardware components separately, and on the system as a whole. Accuracy is measured in spatial and angular deviation and the result is an average accuracy of approximately one degree on synthetic data and 0.24 to 1.5 degrees on real images at a range of 600 mm.

Att applicera ansiktsigenkänning direkt på en mobiltelefon är en utmanande uppgift, inte minst med tanke på den begränsade minnes- och processorkapaciteten samt den stora variationen med avseende på ansiktsuttryck, hållning och ljusförhållande i inmatade bilder. Det är fortfarande långt kvar till ett färdigutvecklat, robust och helautomatiskt ansiktsigenkänningssystem för den här miljön. Men resultaten i det här arbetet visar att genom att plocka ut feature-värden från lokala regioner samt applicera en välgjord warpstrategi för att minska problemen med variationer i position och rotation av huvudet, är det möjligt att uppnå rimliga och användbara igenkänningsnivåer. Speciellt för ett halvautomatiskt system där användaren har sista ordet om vem personen på bilden faktiskt är. Med ett galleri bestående av 85 personer och endast en referensbild per person nådde systemet en igenkänningsgrad på 60% på en svårklassificerad serie testbilder. Totalt 73% av gångerna var den rätta individen inom de fyra främsta gissningarna. Att lägga till extra referensbilder till galleriet höjer igenkänningsgraden rejält, till nästan 75% för helt korrekta gissningar och till 83,5% för topp fyra. Detta visar att en strategi där inmatade bilder läggs till som referensbilder i galleriet efterhand som de identifieras skulle löna sig ordentligt och göra systemet bättre efter hand likt en inlärningsprocess. Detta exjobb belönades med pris för "Bästa industrirelevanta bidrag" vid Svenska sällskapet för automatiserad bildanalys årliga konferens i Lund, 13-14 mars 2008. / Applying face recognition directly on a mobile phone is a challenging proposal due to the unrestrained nature of input images and limitations in memory and processor capabilities. A robust, fully automatic recognition system for this environment is still a far way off. However, results show that using local feature extraction and a warping scheme to reduce pose variation problems, it is possible to capitalize on high error tolerance and reach reasonable recognition rates, especially for a semi-automatic classification system where the user has the final say. With a gallery of 85 individuals and only one gallery image per individual available the system is able to recognize close to 60 % of the faces in a very challenging test set, while the correct individual is in the top four guesses 73% of the time. Adding extra reference images boosts performance to nearly 75% correct recognition and 83.5% in the top four guesses. This suggests a strategy where extra reference images are added one by one after correct classification, mimicking an online learning strategy.

Within this thesis an algorithm for object recognition called Cluster Matching has been developed, implemented and evaluated. The image information is sampled at arbitrary sample points, instead of interest points, and local image features are extracted. These sample points are used as a compact representation of the image data and can quickly be searched for prior known objects. The algorithm is evaluated on a test set of images and the result is surprisingly reliable and time efficient.