Fawnia, Azaria Putri (2026) Implementasi Deep Learning Untuk Sistem Kontrol Perangkat Elektronik Berbasis Gestur Tangan Pada Raspberry Pi 5. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5024221038-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5024221038-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy |
Abstract
Perkembangan Internet of Things (IoT) telah mendorong adopsi smart home yang pesat, namun metode interaksi konvensional seperti aplikasi seluler dan perintah suara masih memiliki keterbatasan. Kontrol berbasis gestur tangan muncul sebagai solusi intuitif dan tanpa kontak fisik, yang semakin relevan dengan kemajuan edge computing seperti Raspberry Pi. Namun, mengimplementasikan pengenalan gestur dinamis berbasis Deep Learning yang kompleks agar berjalan secara real-time pada perangkat edge dengan sumber daya terbatas menghadirkan tantangan komputasi yang signifikan. Penelitian ini bertujuan mengatasi tantangan tersebut melalui pengembangan sistem kontrol terdistribusi yang mengintegrasikan Raspberry Pi 5 sebagai unit pemrosesan Utama dan ESP8266 sebagai aktuator nirkabel. Model Long Short-Term Memory (LSTM) dilatih menggunakan dataset komprehensif yang terdiri dari 10 kelas gestur dinamis untuk menjamin kemampuan generalisasi, yang Kemudian dioptimalkan melalui teknik kuantisasi 8-bit ke format TensorFlow Lite untuk memaksimalkan efisiensi pada perangkat edge. Hasil pengujian empiris menunjukkan bahwa strategi optimasi ini berhasil secara drastic meningkatkan kinerja komputasi, memungkinkan sistem mencapai kecepatan pemrosesan rata-rata 25 FPS dengan latensi inferensi di sisi edge yang sangat rendah di Bawah 30 ms. Meskipun total waktu respons sistem rata-rata 168 ms terbukti sangat responsif untuk interaksi manusia-komputer, analisis mendalam mengungkapkan bahwa komunikasi jaringan nirkabel (WiFi) menjadi hambatan (bottleneck) utama yang mendominasi lebih dari 80% total latensi dan menunjukkan variabilitas (jitter) yang tinggi akibat karakteristik inheren protokol di lingkungan yang rentan interferensi.
================================================================================================================================
The development of the Internet of Things (IoT) has driven the rapid adoption of smart homes, but conventional interaction methods such as mobile applications and voice commands still have limitations. Hand gesture-based control has emerged as an intuitive, contactless solution that is increasingly relevant with the advancement of edge computing such as Raspberry Pi. However, implementing complex Deep Learning-based dynamic gesture recognition to run in real-time on edge devices with limited resources presents significant computational challenges. This research aims to address these challenges through the development of a distributed control system that integrates Raspberry Pi 5 as the main processing unit and ESP8266 as a wireless actuator. A Long Short-Term Memory (LSTM) model was trained using a comprehensive dataset consisting of 10 classes of dynamic gestures to ensure generalization capabilities, which was then optimized through 8-bit quantization techniques into TensorFlow Lite format to maximize efficiency on edge devices. Empirical testing results show that this optimization strategy has dramatically improved computational performance, enabling the system to achieve an average processing speed of 25 FPS with extremely low edge-side inference latency of under 30ms. Although the average total system response time of 168 ms proved to be very responsive for human-computer interaction, in-depth analysis revealed that wireless network communication (WiFi) was the main bottleneck, accounting for more than 80% of the total latency and exhibiting high variability (jitter) due to the inherent characteristics of the protocol in an environment prone to interference.
Actions (login required)
 |
View Item |