Single-Snapshot Pedestrian Gait Recognition at the Edge: A Deep Learning Approach to High-Resolution mmWave Sensing

Taking advantage of recent advancements in the development of high-resolution mmWave sensors, the next generation of autonomous vehicles will incorporate advanced gait analysis deep learning models for precise pedestrian detection in a typical automotive scene. Prior works in pedestrian gait analysis, however, restrict their scope to idealized, controlled environments that neglect many of the cofactors inherent in a practical auto-motive application, while relying upon unlimited processing and computation power to achieve maximal Doppler resolution. The reality is that next-generation, power-efficient sensors deployed at the edge are often resource-limited, and critically require real-time operation to mitigate latency in autonomous decision making, thereby limiting their data processing capacity. In this work, we investigate and quantify the effect of oft-ignored edge system constraints on classifier performance, addressing the use of limited Doppler resolution and the presence of uncontrollable cofactors such as pedestrian location/trajectory and parasitic platform vibration. Results validate the significant impact of such system constraints on classification accuracy, with up to an 11 percent reduction in performance across Doppler resolution and a 40 percent degradation in differentiation between selected gait classes under induced platform vibration.