ABSTRACT Real‐time people counting and indoor positioning are essential features for energy‐aware and privacy‐preserving smart environments. However, achieving low‐latency inference under strict power and bandwidth constraints remains challenging, particularly with conventional high‐resolution RGB or depth‐based vision systems. In this work, we present a fully co‐optimized sensing and inference pipeline that combines a low‐resolution thermopile infrared sensor with a hardware‐tailored detection architecture. A 3232 passive infrared array is used to collect a custom indoor dataset, and a compact pre‐processing chain reduces frame size by over 80% while preserving silhouette information. The resulting binary inputs are processed by a quantization‐ and pruning‐aware YOLOv3‐tiny detector, achieving accurate and compact representation suitable for constrained edge devices. To support deployment on In‐Memory Accelerators (IMC), we experimentally characterize multi‐level cell (MLC) behavior in a fabricated 28 nm FeFET array, including programming voltage characteristics and memory window analysis. Controlled multi‐level current accumulation is further validated at crossbar array level, confirming robust analog dot‐product behavior under realistic operating conditions. The proposed pipeline demonstrates a holistic approach to embedded perception, spanning sensor data capture, model compression, and experimentally validated non‐volatile analog hardware acceleration.
Vardar et al. (Thu,) studied this question.