Real-Time Instance Segmentation for Low-Cost Mobile Robot Systems Based on Computation Offloading

Instance segmentation can enable mobile robots to obtain the environmental semantic information and accomplish more complex interaction with environments, such as navigation, grasping, and virtual reality. However, low-cost mobile robots have limited onboard resources, and can not afford the massive computation of instance segmentation methods. This paper proposes a real-time instance segmentation framework for mobile robot systems based on computation offloading, which offloads part of computation of the instance segmentation network to the cloud, and leverages the powerful computation resources and sufficient memories on the cloud platform to accelerate the network. First, we formulate the instance segmentation network as the directed acyclic graph, and present its time cost model and energy consumption model. Then, a computation offloading strategy is proposed to reduce the time cost of the whole instance segmentation and the energy consumption on the mobile robot. Our framework has been verified on the representative one-stage method, Yolact, and two-stage method, Mask R-CNN. The results show that our framework can accelerate the execution of instance segmentation network on mobile robots, and achieve the speed of around one second per frame.