In-Sensor Motion Recognition with Memristive System and Light Sensing Surfaces
🤖 AI Summary
To address the dual challenges of ultra-low power consumption and high privacy in gesture recognition for edge devices, this paper proposes a photonic-synaptic co-designed sensing-and-computing-in-memory architecture. The method leverages a photosensitive surface to directly capture motion signals and performs analog-domain feature extraction and mapping on-chip; synaptic weights are stored in an HfO₂-based memristor crossbar array, while classification is executed efficiently via a Winner-Take-All circuit. This work introduces the first hardware-level co-design integrating photonic sensing and memristive computing, enabling energy-harvesting operation and fully on-device processing—thereby significantly enhancing both energy efficiency and data privacy. Experimental results demonstrate 97.22% accuracy for four-class gesture recognition under 5% noise, with only 0.952 nJ inference energy and 4.17 nJ weight-mapping energy—two orders of magnitude lower than conventional digital approaches.
Application Category
📝 Abstract
In this paper, we introduce a novel device architecture that merges memristive devices with light-sensing surfaces, for energy-efficient motion recognition at the edge. Our light-sensing surface captures motion data through in-sensor computation. This data is then processed using a memristive system equipped with a HfO2-based synaptic device, coupled with a winner-take-all (WTA) circuit, tailored for low-power motion classification tasks. We validate our end-to-end system using four distinct human hand gestures—left-to-right, right-to-left, bottom-to-top, and top-to-bottom movements—to assess energy efficiency and classification robustness. Our experiments show that the system requires an average of only 4.17 nJ for taking our processed analog signal and mapping weights onto our memristive system and 0.952 nJ for testing per movement class, achieving 97.22% accuracy even under 5% noise interference. A key advantage of our proposed architecture is its low energy requirement, enabling the integration of energy-harvesting solutions such as solar power for sustainable autonomous operation. Additionally, our approach enhances data privacy by processing data locally, reducing the need for external data transmission and storage.
Problem
Research questions and friction points this paper is trying to address.
Energy-efficient motion recognition using memristive and light-sensing systems
Low-power classification of human gestures with high accuracy
Local data processing for enhanced privacy and sustainability
Innovation
Methods, ideas, or system contributions that make the work stand out.
Memristive system with HfO2 synaptic device
Light-sensing surface for in-sensor computation
Winner-take-all circuit for low-power classification
🔎 Similar Papers
No similar papers found.
