Real-Time Mobile Robot Perception Based on Deep Learning Detection Model

Abstract

The recent advances in deep learning models have enabled the robotics community to utilize their potential. The mobile robot domain on which deep learning has the most influence is scene understanding. Scene understanding enables mobile robots to exist and execute their tasks through processes such as object detection, semantic segmentation, or instance segmentation. A perception system that can recognize and locate objects in the scene is of the highest importance for achieving autonomous behavior of robotic systems. Having that in mind, we develop the mobile robot perception system based on deep learning. More precisely, we utilize an accurate and fast Convolution Neural Network (CNN) model to enable a mobile robot to detect objects in its scene in a real-time manner. The integration of two CNN models (SSD and MobileNet) is performed and implemented on mobile robot RAICO (Robot with Artificial Intelligence based COgnition). The experimental results show that the proposed perception system enables a high degree of object recognition with satisfying inference speed, even with limited processing power provided by Nvidia Jetson Nano integrated within RACIO.