Learn to See: A Microwave-based Object Recognition System Using Learning Techniques

Abstract

The capability to recognize nearby objects automatically has numerous applications including asset tracking, lifestyle analysis, and navigation assistance for blind people. In recent years, several approaches were proposed, but they are either limited to electric objects or objects instrumented with tags, which cannot scale. There are also acoustic or vision-based techniques for recognizing uninstrumented objects, but they may have privacy issues. In this paper, we present a microwave-based object detection and recognition approach. Specifically, the proposed system leverages Universal Software Radio Peripherals (USRPs) to transmit microwave signals through the target object and capture them on the opposite side. To reduce the privacy impact, we use a single antenna for receiving a single-pixel “image”. Then, a Random Forest classifier learns the characteristics of the received signals altered by a given object, enabling object recognition. Using a wide range of microwave frequencies, we evaluated the proposed system’s capability to detect and differentiate between four different objects of different materials. The evaluation results show that, using only a signal, the system can correctly detect the presence of the object 98.7% of the time. The system can also differentiate between different objects 92% of the time.