Anomaly Detection Based Image Feature Matching Under Vehicle-Infrastructure Viewpoint

Abstract

Matching local features on two or more images is fundamental and critical for many applications in the field of computer vision. The putative correspondences often contaminated by mismatches when applying feature matching methods in real-world settings. Therefore, the paper focuses on removing mismatches from putative matches by investigating the idea of motion consistency. This paper converts the feature matching into an outlier detection problem based on the converted spatial distance matrix. The classic anomaly detection method LOF is adopted to detect the outliers so that the false matches can be identified and removed. The images for experiments are obtained from cameras mounted on a vehicle and roadside infrastructure. Experimental results demonstrate that the proposed method achieves 88.14% average precision, which is the highest among the other two methods.