Lens-less millimeter-wave (mmWave) imaging of moving objects using a sparse array relies on knowledge of the relative positions between the moving object and the imaging system to enable coherent image reconstruction. However, accurate object position information is rarely available in commercial applications where the moving object, e.g. a conveyor belt or a robot, is controlled independently of the imaging system, or where the imaged objects move autonomously. This poses a significant hurdle for many commercial mmWave imaging applications. We present a video-based motion extraction approach for active mmWave imaging. The object velocity is extracted in real time from motion vectors obtained from a compressed video. This information is combined with readouts from a distance sensor to infer the position of the object at each time instant. Leveraging video-derived motion vectors enables the offloading of computational complexity of 2-D spatial correlations to highly optimized algorithms operating on camera frames. We show experimentally that the image quality of a commercial high-throughput 3-D mmWave imaging system prototype is improved significantly by this approach when the velocity of the target is unknown and time-varying. We furthermore show that image quality is also improved compared to known average motion profiles of the imaged objects. Using a lab setup with known ground truth, we show that the RMS position error is 2.5 mm over a travel length of 0.52 m. This is better than 1/8 of the wavelength at K-band (24 GHz) along the trajectory and thus sufficient to achieve excellent image quality at K-band and longer wavelengths.
|