Continuum Detection and Predictive-Corrective Classification of Crack Networks

Abstract

This paper proposes a crack network detection and classification scheme for perception of highly stochastic road cracks using probabilistic formulations. Contrary to conventional binary detection techniques, the continuum detection approach described here allows features to be extracted from crack images with allowance for uncertainty in detection. Furthermore, multi-dimensional prediction and belief fusion in the feature space is afforded by the sequential nature of data collection; classification is then carried out by probabilistic decision-making. These methods have shown superior performance to simplistic conventional approaches, offering as much as a 14% increase in accuracy when compared to naïve classification, even without any parameter optimization. Furthermore, predictive-corrective classification yields a 28% increase in variance of class probability assignment; this means that, in addition to improved classification results, correct classes are assigned with greater confidence when compared to simpler methods.