Deep learning based smart radar vision system for object recognition

The capability of manipulating objects based on vision system are essential to robotic applications. For a vision system, high accuracy and robustness is vital. However, due to external environment and interference influences, as well as to lighting conditions influences, the most existing image-based vision system are limited in applications. In this context, a novel and robust data-driven vision system for recognizing different types of objects in various external situations has been presented in this paper. The system consists of three parts, namely, integrated module based on X2 Ultra-wideband (UWB) radar, deep learning (DL) module and robot controller. Specifically, the introduction of X2 UWB radar aims to (1) capture and process radar echo signals of different objects in various environments; (2) extract Range-Doppler power spectrograms from the processed signals accurately. The second part combines the DL algorithm to take the spectrograms as the input from the former to realize object classification. And the last part is used to control the behavior of the robot based on the classification results. The experimental results show that the proposed vision system achieves very reliable performance in recognizing six types of objects in four kinds of external environment and lighting conditions with three deep learning network models. Besides, by experimental data comparison, the performance of the proposed system achieves higher than that of the image-based system in the same poor conditions and ResNet is the optimal model for our proposed system in all conditions.