| dc.description.abstract | The past few years have witnessed growing interest in using millimeter-wave signals for non-line-of-sight (NLOS) perception tasks, with applications in robotics, augmented reality, and smart-homes. However, existing systems suffer from a lack of large mmWave datasets, resulting in limited accuracy and generalizability compared to their line-of-sight, camera-based counterparts. We present the design, implementation, and evaluation of mmSim, a new, high-speed millimeter-wave (mmWave) simulator capable of producing large synthetic datasets to help drive the field of mmWave-based NLOS perception. mmSim introduces two main contributions to improve the speed over existing mmWave simulators. First, it pre-selects areas of the object, which will produce reflections towards each simulated antenna location, allowing it to minimize future computation. Second, it introduces a coarse-to-fine approach to allow early, less critical steps to operate at lower resolutions, while maintaining the high resolution in later steps required for high-accuracy images. These techniques, combined with other performance optimizations, allow mmSim to achieve an over 24x improvement in speed over state-of-the-art mmWave simulators. | |
