High Clutter, Close Range, Wi-Fi Imaging and Probabilistic, Learning Classifier (Poster)

Abstract

While robotics have been used in manufacturing for decades, we are now seeing robots whether animal-like or human-like open doors and maneuver indoors. To maneuver indoors these robots need to be aware of their situation within each room and the objects within. It’s no longer acceptable for a robot to use just pulses for collision avoidance; it must interact with static (non-moving) objects, too. Today we have a large array of imaging methods being visible light imaging, thermal imaging, night-vision imaging, but we need imaging in other cases such as a smoke-filled area containing same temperature objects. To see static objects in this arena, this research introduces Wi-Fi static object imaging and classification. It will allow a robot to maneuver a room with static objects when these other imaging methods won’t work well. This involves many challenges. The images created will be barely identifiable due to the low resolution of Wi-Fi, but the images need to be classified (identified). There are two major portions to this research. The first is the signal processing portion, where images are created, and the second is the image classification portion. In the signal processing portion, images will be created by using Wi-Fi signals transmitted and received on Ettus Universal Software Radio Peripheral (USRP) hardware and directional antennas. These USRP’s are interfaced to GNU Radio for which the researchers have developed specialized routines to implement this portion of the research. In the image classification phase, the images created will be very blob-like with different reflective intensities. These images need some form of intelligence to classify them, and the system needs to improve its classification over time. AI neural networks will be employed and developed to work on the images. This research is a work-in-progress where the signal processing portion is complete except for further tuning. The images created so far have very distinct characteristics. This means these Wi-Fi images will be good candidates for the classification phase.