Rapid cascade condition assessment of ductwork via robot vision

Abstract

Automatic assessment of condition in ductwork is very desirable in applications. Presented is a visual condition diagnosis approach, which is capable of processing images rapidly and achieving high accuracy rates. A hierarchical coarse-to-fine image segmentation method is employed. False alarms could thus be progressively eliminated, which is robust in strongly noisy conditions. The simple classifiers combined in a cascade quickly classify the detected images and discard the uninterested (non-object) images, leaving more computation power on promising object-like regions. The features of each simple classifier are selected based on the Bhattacharyya distance. The cascade can be viewed as an object-specific focus-of-attention mechanism. Experimental results validate the effectiveness and rapidity of the proposed assessment method.