融合空洞卷积神经网络的语义SLAM研究

Research on semantic SLAM of fusion dilated convolutional neural network

为了解决传统视觉SLAM算法受动态环境因素影响较大、对设备的算力要求高的问题,该文提出一种融合ORB-SLAM2、语义标签以及全局性稠密光流法的视觉SLAM算法。该方法采用空洞卷积神经网络的语义分割模块为双目图像添加语义标签,识别物体类别。再结合相邻帧间位置信息对图像的动态点进行剔除。最后使用Octo-map优化定位与建图,实现动态环境下地图的建立与实时更新。实验结果证明,该文提出的算法在动态环境下的特征点提取速度和质量相较传统视觉SLAM算法有一定提高。

A visual SLAM algorithm fusing ORB-SLAM2,semantic label and global dense optical flow method is proposed to solve the problem that the traditional visual SLAM algorithm is greatly affected by dynamic environment factors and requires high computing power of equipment.In this method,the semantic segmentation module of dilated convolution neural network is used to add semantic tags to binocular images,so as to identify object categories.The dynamic points of the image are eliminated by combining the position information between adjacent frames.The octo-map is used to optimize location and map construction to realize the map establishment and real-time update under the dynamic environment.The experimental results show that,in combination with the traditional SLAM algorithm,the speed and quality of feature point extraction in dynamic environment are improved.