基于SLAM算法和深度神经网络的语义地图构建研究

RESEARCH ON SEMANTIC MAPPING BASED ON SLAM ALGORITHM AND DEEP NERUAL NETWORK

语义地图在移动机器人的导航等任务中有着关键的作用。目前基于视觉的机器人自动定位和制图(SLAM)系统已经能够达到较高的定位精度要求,但是基于多目几何的视觉SLAM算法并未充分利用空间中丰富的语义信息,地图中保留的地图点信息只是没有语义的空间几何点。由于基于卷积神经网络的算法在目标检测领域取得了突破性的成绩,利用目前最新的基于卷积神经网络的目标检测算法YOLO,实现场景的实时目标检测,并结合SLAM算法构建语义地图。将视觉SLAM定位的精确性和深度神经网络在语义提取方面的优势相结合,既能够提高SLAM算法的准确性,同时采集到的丰富图像信息又能进一步训练更加深的神经网络。该算法能够应用于机器人的智能导航,同时也能作为数据采集器为其他视觉任务提供具备语义与几何信息的图像数据。

Semantic mapping plays a key role in the task of mobile robot navigation. At present, the vision-based SLAM system has been able to achieve higher accuracy requirements, but the visual SLAM algorithm based on multi-view geometry cannot use the rich semantic information in space. The map point information reserved in the map is only spatial geometric point without semantic information. Since the algorithm based on convolution neural network has made a breakthrough in the field of object detection, the semantic mapping is constructed using the latest YOLO algorithm of object detection based on convolution neural network, which realized the real-time object detection in scene, combined with the SLAM algorithm. The proposed approach combines the accuracy of visual SLAM and the advantages of deep neural network in semantic extraction, which can improve the accuracy of SLAM algorithm, and the rich images collected can further help us to train deeper neural networks. This algorithm can be applied to the intelligent navigation of the robot, but also as a data collector for other visual tasks to collect the semantic and geometric information with the image data.