Multi-scale matching for data association in vision-based SLAM

In this paper, we propose a multi-scale matching approach to address the data association problem in vision-based simultaneous localization and mapping (SLAM). Data association in vision-based SLAM can be simply represented as a feature correspondence problem related to two features observed in different positions under different imaging conditions. We apply an improved Harris detector to automatically extract feature points with high localization accuracy. The scale space in frequency domain is built by introducing the Log-Gabor filter under the monogenic signal analysis framework. Reliable correspondence between two features is found and identified over all scales by combining advantages of geometric invariant property in monogenic signal information as well as photometric invariant property in color entropy information. Our approach is able to establish correct data association which is robust to changes in scale, blur, viewpoint, and illumination. Moreover, the cost on map management is reduced by selecting the obtained small number of reliably matched features as visual landmarks. Experiments conducted on a standard benchmark dataset and an office-like indoor environment demonstrate the effectiveness of our approach.