End-to-End Vision-to-Motion Model with Auxiliary Segmentation Module for Indoor Navigation

The majority of vision-based navigation algorithms consist of two independent sub-algorithms, environment perception algorithm and motion planning algorithm. Environment perception involves the positioning of the navigation robot and the construction of obstacle avoidance maps, from which motion decisions are drawn with planning algorithms. This type of algorithm requires special equipment such as laser camera and the hierarchical design tend to amplify errors transferred between upstream and downstream algorithms. To tackle the problems above, an end-to-end model based on deep learning is proposed to achieve a direct mapping from vision to motion under unconstructed indoor environment. To implicitly guide the mode to learn an internal ability of obstacle avoidance with the prior map information, an auxiliary semantic segmentation module is introduced into the backbone model, yielding a multi-task loss for the optimization of the model. Eventually, transfer learning technique is applied to extend training dataset with an opensource dataset and strengthen the generalization capability of the model. Experimental results demonstrate that the proposed framework is accurate, efficient and feasible in the real application.