Scalable Non-Parametric Parsing for Segmentation and Recognition of High-Quantity Low-Cost Highway Assets from Car-Mounted Video Streams

摘要

Systematic condition assessment on high-quantity low-cost highway assets requires frequent reporting on location and up-to-date status of these assets. Recent research on video-based assessment of assets have primarily focused on detecting traffic signs during data collection, and are not applicable to other assets such as guardrails and light poles. To overcome such limitations, this paper presents fast graph-based segmentation and super-parsing algorithms which efficiently segment highway assets from 2D video streams. Using a fast graph-based segmentation algorithm, superpixels are obtained from each frame and their appearance is computed using a histogram of textons and dense SIFT-descriptors. A likelihood ratio score is obtained for each superpixel and an asset label that maximizes the ratio is assigned. Given a frame to be interpreted, the algorithm performs global matching against the training set, followed by superpixel-level matching and efficient Markov Random Field (MRF) optimization. The MRF simultaneously labels video frame regions into semantic and geometric classes of assets. Experimental results are presented on the Virginia Tech Smart Road research facility in a 2.2 mile highway. The work contributes to the body-of-knowledge by detecting 3D assets that have not previously been detectable by the state-of-the-art methods. It also enables further development of techniques that can recognize subcategories of highway assets.