Perkembangan teknologi kecerdasan buatan, khususnya pada bidang computer vision,
telah mendorong pengembangan sistem deteksi objek secara near real-time yang dapat
dimanfaatkan sebagai perangkat bantu bagi penyandang tunanetra. Penelitian ini
bertujuan untuk merancang dan membangun sistem kalung pintar berbasis Raspberry Pi
4 yang mampu mendeteksi objek di lingkungan sekitar pengguna menggunakan metode
You Only Look Once (YOLOv8), dilengkapi sensor ultrasonik HC-SR04 untuk mengukur
jarak halangan, output suara berbasis text-to-speech untuk menyampaikan informasi,
serta modul GPS BN-880 yang terintegrasi dengan Telegram Bot untuk pengiriman
informasi lokasi pengguna melalui jaringan internet. Pengujian sistem dilakukan terhadap
lima subsistem, yaitu deteksi objek, sensor ultrasonik, output suara, GPS dan Telegram,
serta pengujian keseluruhan sistem. Hasil pengujian menunjukkan sistem deteksi objek
mencapai tingkat Keberhasilan deteksi sebesar 85,29% dengan Keberhasilan deteksi lebih
tinggi pada kondisi pencahayaan terang (94,59%) dibandingkan kondisi redup (74,19%).
Sensor ultrasonik menghasilkan rata-rata error sebesar 5,42% (akurasi 94,58%), dengan
performa optimal pada rentang 50–200 cm namun menurun pada jarak di atas 250 cm.
Modul output suara berhasil menyampaikan informasi dengan tingkat kesesuaian 100%
dan rata-rata delay 0,739 detik. Modul GPS berhasil mengirimkan lokasi ke Telegram
pada seluruh pengujian (100%), meskipun rata-rata Time-To-First-Fix (TTFF) sebesar
185,56 detik masih jauh melebihi spesifikasi datasheet. Pengujian keseluruhan sistem
pada tiga kondisi lingkungan (normal, abnormal, dan malam hari) menghasilkan tingkat
keberhasilan sebesar 95,83%. Hasil pengujian menunjukkan sistem Kalung AI mampu
mendeteksi objek dan memberikan informasi melalui output suara dengan rata-rata delay
sebesar 0,739 detik dan delay maksimum 2,51 detik. Hasil ini menunjukkan bahwa sistem
deteksi objek pada Kalung AI mampu beroperasi secara near real-time, sedangkan Voice
Assistant masih memiliki keterbatasan dalam waktu respons.
The development of artificial intelligence technology, particularly in the field of computer
vision, has encouraged the development of near real-time object detection systems that
can be utilized as assistive devices for visually impaired individuals. This research aims to
design and build a smart necklace system based on Raspberry Pi 4 that is capable of
detecting objects in the user's surrounding environment using the You Only Look Once
(YOLOv8) method, equipped with an HC-SR04 ultrasonic sensor to measure obstacle
distance, a text-to-speech-based audio output to convey information, and a BN-880 GPS
module integrated with a Telegram Bot to transmit the user's location through an internet
connection. System testing was conducted on five subsystems, namely object detection,
the ultrasonic sensor, audio output, GPS and Telegram, as well as testing of the overall
system. The test results show that the object detection system achieved a detection
success rate of 85.29%, with a higher success rate under bright lighting conditions (94.59%)
compared to dim conditions (74.19%). The ultrasonic sensor produced an average error of
5.42% (94.58% accuracy), with optimal performance in the 50–200 cm range but
decreasing performance beyond 250 cm. The audio output module successfully delivered
information with a 100% conformity rate and an average delay of 0.739 seconds. The GPS
module successfully transmitted location data to Telegram in all trials (100%), although the
average Time-To-First-Fix (TTFF) of 185.56 seconds still far exceeded the datasheet
specification. Overall system testing under three environmental conditions (normal,
abnormal, and nighttime) achieved a success rate of 95.83%. The results indicate that the
Kalung AI system is able to detect objects and deliver information through audio output with
an average delay of 0.739 seconds and a maximum delay of 2.51 seconds. These findings
show that the object detection system in Kalung AI is able to operate in near real-time, while
the Voice Assistant still has limitations in response time.