Systematic condition assessment of high-quantity low-cost roadway assets such as traffic signs, guardrails, and pavement markings requires frequent reporting on location and up-to-date status of these assets. Today, most Departments of Transportation (DOTs) in the US collect data using camera-mounted vehicles to filter, annotate, organize, and present the data necessary for these assessments. However, the cost and complexity of the collection, analysis, and reporting as-is conditions result in sparse and infrequent monitoring. Thus, some of the gains in efficiency are consumed by monitoring costs. This dissertation proposes to improve frequency, detail, and applicability of image-based condition assessment via automating detection, classification, and 3D localization of multiple types of high-quantity low-cost roadway assets using both images collected by the DOTs and online databases such Google Street View Images. To address the new requirements of US Federal Highway Administration (FHWA), a new method is also developed that simulates nighttime visibility of traffic signs from images taken during daytime and measures their retro-reflectivity condition.To initiate detection and classification of high-quantity low-cost roadway assets from street-level images, a number of algorithms are proposed that automatically segment and localize high-level asset categories in 3D. The first set of algorithms focus on the task of detecting and segmenting assets at high-level categories. More specifically, a method based on Semantic Texton Forest classifiers, segments each geo-registered 2D video frame at the pixel-level based on shape, texture, and color. A Structure from Motion (SfM) procedure reconstructs the road and its assets in 3D. Next, a voting scheme assigns the most observed asset category to each point in 3D. The experimental results from application of this method are promising, nevertheless because this method relies on using supervised ground-truth pixel labels for training purposes, scaling it to various types of assets is challenging. To address this issue, a non-parametric image parsing method is proposed that leverages lazy learning scheme for segmentation and recognition of roadway assets. The semi-supervised technique used in the proposed method does not need training and provides ground truth data in a more efficient manner. It is easily scalable to thousands of video frames captured during data collection. Once the high-level asset categories are detected, specific techniques needs to be exploited to detect and classify the assets at a higher level of granularity. To this end, performance of three computer vision algorithms are evaluated for classification of traffic signs in presence of cluttered backgrounds and static and dynamic occlusions. Without making any prior assumptions about the location of traffic signs in 2D, the best performing method uses histograms of oriented gradients and color together with multiple one-vs-all Support Vector Machines, and classifies these assets into warning, regulatory, stop, and yield sign categories. To minimize the reliance on visual data collected by the DOTs and improve frequency and applicability of condition assessment, a new end-to-end procedure is presented that applies the above algorithms and creates comprehensive inventory of traffic signs using Google Street View images. By processing images extracted using Google Street View API and discriminative classification scores from all images that see a sign, the most probable 3D location of each traffic sign is derived and is shown on the Google Earth using a dynamic heat map. A data card containing information about location, type, and condition of each detected traffic sign is also created. Finally, a computer vision-based algorithm is proposed that measures retro-reflectivity of traffic signs during daytime using a vehicle mounted device. The algorithm simulates nighttime visibility of traffic signs from images taken during daytime and measures their retro-reflectivity. The technique is faster, cheaper, and safer compared to the state-of-the-art as it neither requires nighttime operation nor requires manual sign inspection. It also satisfies measurement guidelines set forth by FHWA both in terms of granularity and accuracy.To validate the techniques, new detailed video datasets and their ground-truth were generated from 2.2-mile smart road research facility and two interstate highways in the US. The comprehensive dataset contains over 11,000 annotated U.S. traffic sign images and exhibits large variations in sign pose, scale, background, illumination, and occlusion conditions. The performance of all algorithms were examined using these datasets. For retro-reflectivity measurement of traffic signs, experiments were conducted at different times of day and for different distances. Results were compared with a method recommended by ASTM standards. The experimental results show promise in scalability of these methods to reduce the time and effort required for developing road inventories, especially for those assets such as guardrails and traffic lights that are not typically considered in 2D asset recognition methods and also multiple categories of traffic signs. The applicability of Google Street View Images for inventory management purposes and also the technique for retro-reflectivity measurement during daytime demonstrate strong potential in lowering inspection costs and improving safety in practical applications.
【 预 览 】
附件列表
Files
Size
Format
View
Image-based recognition, 3D localization, and retro-reflectivity evaluation of high-quantity low-cost roadway assets for enhanced condition assessment
PDF
download
878987797
028-85220240
