Robust Candidate Pruning Approach Based on the Dempster-shafer Evidence Theory for Fast Corner Detection with Noise Tolerance in Gray-level Images

A fast two-stage corner detector with noise tolerance is presented in this paper. At the first stage, candidate-corner pixels are selected by a proposed candidate pruning approach. At the second stage, real corners are recognized by the Harris detector among the candidate-corner pixels. In general, corners are considered as the junction of edges. Therefore, edge pixels with a high gradient in more than one direction can be selected as candidate-corner pixels. Meanwhile, noisy pixels always cause false detections in most corner detectors. Those noisy pixels thus must be excluded from candidate-corner pixels to enhance the noise tolerance capability. In this paper, candidate-corner pixels are selected based on local features that are extracted from a sliding observation window. These features including gradient, edge and impulse noise are regarded as pieces of evidence and are further combined by Dempster’s rule to yield a final belief value. Through the well-selection of candidate-corner pixels, the candidate pruning approach can 1) enhance the noise tolerance capability, and 2) reduce the computational cost of the proposed corner detector. Experimental results show that the proposed method outperformed other well-known corner detectors.