MobileNet-DeepLabV3+ based Robot Passable Path Segmentation and Navigation Line Extraction

Abstract

Traditional vision-based navigation methods for mobile robot can only be applied to some specific simple scenes due to the limitations of complex road conditions and light variations. However, the research achievements of deep learning and the substantial improvement in computer data processing performance have made it practical and applicable to guide robots in navigation using visual semantic information. For the robot visual semantic navigation problem, a lightweight DeepLabV3+ image semantic segmentation model based on MobileNetV2 is constructed as the backbone network. The images acquired by the robot vision sensor can be sensed and segmented into targets and feasible paths within the scene. Navigation lines for navigable paths can be established using the obtained feasible areas. In this way, the robot's motion decisions can be guided. Therefore, the mobile robot is capable of pure vision-based navigation. The segmentation algorithm is validated using the CamVid dataset and the constructed experimental environment dataset with mPA 82.73 and mIoU 71.8, and the navigation line can be generated according to the drivable path, which guides the robot's navigation.