Yakınlık Sensör Dizilerinde Yapay Zeka Yongası Tabanlı Hareket Algılama ve Tanıma

Abstract

In this thesis, the stages of designing an artificial intelligence chip capable of recognizing hand gestures are presented, utilizing the trained weight and bias values derived from data generated by capacitive proximity sensor (CPS) arrays. Numerous systems for motion detection and recognition exist in the literature. The study aims to fulfill various software and hardware requirements for motion detection and recognition using a single integrated circuit, specifically an artificial intelligence (AI) chip. A dataset collected using CPS arrays was utilized in the thesis study. 75% of the data obtained from hand gestures was set aside as training, 15% as validation, and 10% as test data. For motion detection and recognition, an AI chip was designed to run a Multi-Layer Perceptron (MLP) architecture, which consists of 4 layers, including input and output layers. Each layer has 36, 20, 20 and 4 nodes, respectively. MLP architecture achieved 99.58% training accuracy and 99.50% F1 score in a computer-based simulation environment. According to the analysis results obtained from the study, the basic building blocks of the architecture to be implemented in the digital design were determined. The digital design was created based on the algorithm tested in the computer environment, and at the final stage, an applicationspecific AI chip was designed. The digital design used a 16-bit fixed-point number format. AI chip, designed using the RTL2GDSII approach with the application-specific integrated circuit (ASIC) concept, achieved a 4 times higher frequency than the design made using a commercial FPGA SoC while consuming 7.25 times less power. Within the scope of the thesis study, a design was realized for the AI chip concept, and a method was presented for the design flow process of the system and subsystems containing the artificial intelligence chip for a specific problem.