基于多特征融合的高鲁棒性视觉SLAM改进算法

An improved algorithm of high robust visual SLAM based on multi-feature fusion

特征提取的有效性对视觉SLAM的性能具有重要影响。为了提高匹配精度,本文提出了一种基于改进oAGAST算法和旋转rLATCH特征(OARL)的二进制字符串描述局部图像的新方法。首先,在图像金字塔尺度空间上进行自适应通用加速分割(AGAST)检测;然后,采用灰度质心方法来进行方向补偿;最后,使用可学习的3个图像块的LATCH特征用于生成特征点描述符。在求取特征点主方向阶段太依赖像素邻域灰度质心法,提出一种改进的匹配算法,采用汉明距离匹配和余弦相似度相结合的方法进行特征点匹配。本文在特征点提取方法以视觉SLAM的ORB算法和其他算法为对照,分别进行不同视角、光照和尺度的识别实验。实验结果显示所提出的OARL算法较ORB算法将精度提高了5%以上,但速度却依旧可以达到实时运算。

The effectiveness of the feature extraction has an important influence on the performance of the visual SLAM.This paper proposes a new method for describing local images based on binary strings with Oriented AGAST and Rotated LATCH(OARL)method to improve the accuracy of visual SLAM.Firstly,Adaptive and Generic Corner Detection based on the Accelerated Segment Test(AGAST)is adopted to detect corner feature on the scale space of image pyramid.Then,the intensity centroid method is used to obtain orientation compensation.Finally,the Learned Arrangements of Three Patch Codes is used to describe the feature.An improved algorithm is proposed in which a combined measure of distance similarity matching with cosine similarity matching is adopted to the main direction of the feature point,while it is dependent on the pixel neighborhood’s intensity centroid method.In this paper,the feature detection is based on the visual SLAM ORB algorithm and other algorithms,and the different perspective,illumination and scale recognition experiments are carried out.The experimental results show that the proposed OARL algorithm improves the accuracy by more than5%compared with the ORB algorithm,also the speed can still achieve real-time operation.