HW-SW Optimization of DNNs for Privacy-preserving People Counting on Low-resolution Infrared Arrays

TL;DR

This paper presents an automated optimization framework for DNNs tailored for privacy-preserving people counting using low-resolution IR sensors, achieving significant reductions in model size, code size, and energy consumption.

Contribution

It introduces a comprehensive full-stack optimization process combining neural architecture search, quantization, and hardware design for efficient DNN deployment.

Findings

Achieved up to 4.2x model size reduction.

Realized 23.8x code size reduction.

Attained 15.38x energy reduction at the same accuracy.

Abstract

Low-resolution infrared (IR) array sensors enable people counting applications such as monitoring the occupancy of spaces and people flows while preserving privacy and minimizing energy consumption. Deep Neural Networks (DNNs) have been shown to be well-suited to process these sensor data in an accurate and efficient manner. Nevertheless, the space of DNNs' architectures is huge and its manual exploration is burdensome and often leads to sub-optimal solutions. To overcome this problem, in this work, we propose a highly automated full-stack optimization flow for DNNs that goes from neural architecture search, mixed-precision quantization, and post-processing, down to the realization of a new smart sensor prototype, including a Microcontroller with a customized instruction set. Integrating these cross-layer optimizations, we obtain a large set of Pareto-optimal solutions in the 3D-space of energy, memory, and accuracy. Deploying such solutions on our hardware platform, we improve the state-of-the-art achieving up to 4.2x model size reduction, 23.8x code size reduction, and 15.38x energy reduction at iso-accuracy.