BP神经网络在SLAM特征匹配中的应用

Application of BP neural network in SLAM feature matching

针对移动机器人进行SLAM特征点提取时,因底层机器人搭载的硬件性能低、带宽窄,使用特征点提取算法对图像进行处理时要花费大量时间这一问题,该文提出一种新的方法。首先利用BP神经网络对图像进行压缩,然后进行特征提取,最后再通过RANSAC算法剔除误匹配。通过实验,在使用不同的压缩参数进行图像压缩重建后,再使用SIFT、SURF、ORB、AKAZE和BRISK 5种算法,在旋转、比例变化、模糊、视角变换、光照和JPEG压缩等情况下,均有良好的匹配效果,压缩后图像大小大大减小,并保证了图像的质量,与原处理方法相比,匹配的准确度和匹配时间均优于原方法。

When the mobile robot carries out simultaneous localization and mapping(SLAM) feature point extraction,due to the low performance and narrow band width of the hardware carried by the underlying robot,it takes a lot of time to process the image with the feature point extraction algorithm. To solve this problem,a new method was proposed in paper. Firstly,back propagation(BP) neural network was used to compress the image,then feature extraction was carried out,and finally the wrong matching was eliminated by random sample consensus(RANSAC) algorithm. Through experiments,after image compression and reconstruction using different compression parameters,SIFT,SURF,ORB,AKAZE and BRISK algorithms had good matching effect on the conditions of rotation,proportion change,blur,angle change,light and JPEG compression. The size of the compressed image was greatly reduced and the quality of the image was guaranteed. Compared with the original method,the matching accuracy and matching time were better than the original method.