Feature-based level of service classification for traffic surveillance

A novel level of service (LOS) estimation approach based on the extraction of three local visual features is presented. The feature set comprises KLT motion vectors and Sobel edges, and is fed into a Gaussian radial-basis-function (GRBF) network to classify the prevailing LOS. The whole approach is designed and implemented to run on smart cameras in real-time and has been evaluated with a comprehensive set of real-world training and test video data from a national motorway. The evaluations in daylight environments have shown an average accuracy of LOS classification of 86.2% on an Atom-based smart camera, with a maximum reachable processing frame rate of 12.5 frames/sec. Incorrect classified samples differed from the ground truth by only one level. The comparisons are done with observation data from sensors utilizing a combination of Doppler radar, ultrasound, and passive infrared technologies.