Object detection based on scale-invariant partial shape matching----Shenyang Institute Of Automation ,Chinese Academy Of Sciences

a Corresponding parts of model and image edges obtained by our method. B Detection results of our method

For human perception, shape alone can often provide sufficient information for successful generic object detection classification, shape retrieval, etc. Moreover, shape is invariant to lighting conditions and is relative stable compared to intra-category appearance variations, which makes it a powerful cue for object detection.

To achieve the purpose of object detection in computer vision, shape descriptor or representation is the key technology for shape-based object detection, which can be classified as global shape descriptor and local/partial shape descriptor.

Global shape descriptors provide good performance on simple scene images or images in which object boundaries can be completely extracted. But their performance decreases seriously when local shape distortion and occlusion occur. Partial shape descriptors can somewhat overcome this issue for their capability of handling occlusion problem in the complex scene and allowing partial matching. However, partial shape matching is always challenged by the following problems: (1) inconsistent fragment extraction, (2) scale and rotation issues. For inconsistent fragment extraction, it is hard to extract identical fragment from the same part of object in different images, e.g. different length or starting position extraction caused by unstable edge detection, clutter background, or occlusion.

The main contributions of this paper are three folds: (1) We propose a partial shape matching method by selecting the corresponding parts of two matching fragments to overcome the inconsistency of partial shape extraction. (2) Local scales for each pair of corresponding points can be calculated, and they play an important role in the matching and detection stages. (3) A two-step detection approach is designed to get the accurate object boundaries: a weighted hough voting method for candidate object location and a point-to-point fine matching process for precise edge location.

Click the following link to see the paper

http://www.sciencedirect.com/science/article/pii/S0006291X15308263