Real-Time Object Recognition Based on NAO Humanoid Robot

Abstract

This paper focuses on the real-time object recognition based indoor humanoid robots like Nao robots. Improving the perceptive ability of service robot has always been a research hotspot. The breakthrough of computer vision technology represented by object recognition provides a broader idea for this purpose. We deployed a micro-cloud layer that connects the robot with the computer vision, thereby realized the concepts of RaaS (Robot as a service). In this paper, in order to make the Nao robot to detect objects faster. We present an architecture about real-time object recognition on Nao, and offload the task of control and data collection from robot to a PC. Next, the image data is transmitted over Ethernet to the workstation, which runs multiple parallel image processing services. These services are built with the current popular deep neural network by TensorFlow and running on a GPU GTX1080 Ti. In the micro-cloud layer, we designed a universal robotic visual task queue model, and a PC registers the task queue to the LAN. There are multiple workers in the LAN, and each worker is an independent service processer. Service processer obtains the task queue from the network and processes the queue, and then the processer puts the results back to the manager. The experimental results of the Nao robot in the simulation and real word show that our model and method are effective. The robot can recognize about 90 kinds of common objects, and each frame of image processing time is about 100 milliseconds.