Real-Time Instance-Aware Segmentation and Semantic Mapping on Edge Devices

Perceiving the environment semantically in real-time is challenging for unmanned aerial vehicles (UAVs) with limited computational resources. In this article, a real-time instance-aware segmentation and semantic mapping method on small edge devices is proposed. Taking red, green, blue, and the depth (RGB-D) image as input, the presented instance segmentation pipeline is able to run at the speed of 38 frames/s on AGX Xavier. To achieve this, we take a lightweight object detection model as the backbone and reformulate the mask generation problem as threshold regression in depth by a novel designed truncation network. After that, a probability grid map is constructed to integrate the categories of voxels and object-level entities. Objects parameterized by pose, extent, category, and point cloud are tracked and fused across frames by data association. Finally, autonomous exploration experiments of UAVs are conducted to demonstrate the effectiveness of the proposed method in both simulation and real-world.

Automated summary

In this article, a real-time instance-aware segmentation and semantic mapping method for UAVs on small edge devices is proposed. By using a lightweight object detection model as the backbone and reformulating the mask generation problem as threshold regression in depth by a novel designed truncation network, the proposed method achieves a speed of 38 frames/s. Autonomous exploration experiments of UAVs demonstrate the effectiveness of the method in both simulation and real-world.