A Fast Image Matching Approach for Real-Time Embedded Systems

With the development of real-time embedded systems, image matching based on binary feature descriptors is getting more and more attention for the needs of devices' low computation complexity and lower memory. However, the main approaches such as BRISK and ORB cannot well juggle the speed and robustness. In this paper, a special and efficient image matching approach for real-time embedded system based on combined algorithms was proposed. For the target of faster performance, CenSurE algorithm adopting STAR filter instead of BLoG filter was chosen as features detector and FREAK algorithm using Mean filter to replace Gaussian filter was introduced in features description. The integral image processing method also plays important roles in operation. Moreover, faster matching strategy named multiple hierarchical clustering trees was applied to deal with massive binary features matching. The comparative experiments for viewpoint, blur, illumination, rotation and scale indicate that our image matching approach possesses excellent speed and robustness.