A new local-feature framework for scale-invariant detection of partially occluded objects

Partially occluded objects are typically detected using local features (also known as interest points, keypoints, etc.). The major problem of the local-feature approach is the scale-invariance. If the objects have to be detected in arbitrary scales, either computationally complex methods of scale-space (multi-scale approach) are used, or the actual scale is estimated using additional mechanisms. The paper proposes a new type of local features (keypoints) that can be used for scale-invariant detection of known objects in analyzed images. Keypoints are defined as locations at which selected moment-based parameters are consistent over a wide radius of circular patches around the keypoint. Although the database of known objects is built using the multi-scale approach, analyzed images are processed using only a single-scale. The paper focuses on the keypoint building and matching only. Higher-level issues of hypotheses building and verification (regarding the presence of known objects) are only briefly mentioned.