Only vision-based navigation is the key of cost reduction and widespread application of indoor mobile robot. Consider the unpredictable nature of artificial environments, deep learning techniques can be used to perform navigation with its strong ability to abstract image features. In this paper, we proposed a low-cost way of only vision-based perception to realize indoor mobile robot navigation, converting the problem of visual navigation to scene classification. Existing related research based on deep scene classification network has lower accuracy and brings more computational burden. Additionally, the navigation system has not yet been fully assessed in the previous work. Therefore, we designed a shallow convolutional neural network (CNN) with higher scene classification accuracy and efficiency to process images captured by a monocular camera. Besides, we proposed an adaptive weighted control (AWC) algorithm and combined with regular control (RC) to improve the robot’s motion performance. We demonstrated the capability and robustness of the proposed navigation method by performing extensive experiments in both static and dynamic unknown environments. The qualitative and quantitative results showed that the system performs better compared to previous related work in unknown environments.

只有基于视觉的导航才是降低成本和室内移动机器人广泛应用的关键。考虑到人工环境的不可预测性，深度学习技术凭借其强大的抽象图像功能，可用于执行导航。在本文中，我们提出了一种基于视觉的低成本感知方法来实现室内移动机器人导航，将视觉导航问题转化为场景分类。现有基于深度场景分类网络的相关研究精度较低，带来较大的计算负担。此外，在先前的工作中尚未对导航系统进行全面评估。因此，我们设计了一种具有较高场景分类精度和效率的浅层卷积神经网络（CNN），以处理单眼相机捕获的图像。此外，我们提出了一种自适应加权控制（AWC）算法，并与常规控制（RC）相结合，以提高机器人的运动性能。通过在静态和动态未知环境中进行广泛的实验，我们证明了所提出的导航方法的功能和鲁棒性。定性和定量结果表明，该系统在未知环境中的性能优于以前的相关工作。