Electronic Theses and Dissertation

Mixed-traffic adalah salah satu jenis atau model lalu lintas yang melibatkan berbagai jenis kendaraaan yang saling berbagi infrastruktur jalan yang sama tanpa adanya pembatas atau pemisah fisik. Lingkungan lalu lintas semacam ini umumnya ditemui di daerah perkotaan di negara-negara berkembang seperti Tiongkok, India, dan Indonesia. Dalam lingkungan lalu lintas campuran, terdapat berbagai jenis objek yang beragam yang perlu diidentifikasi lebih lanjut. Kompleksitas lingkungan ini mengharuskan deteksi objek yang akurat. Namun deteksi objek yang digunakan belum optimal. Visibilitas dalam kondisi lalu lintas campuran seringkali rendah, objek-objek dalam lingkungan bergerak ke arah yang sama atau berlawanan, yang semakin meningkatkan kompleksitas skenario lalu lintas di suatu jalan tertentu. Tujuan penelitian ini adalah untuk mengevaluasi kinerja antara model pretrained dan trained berbasis Convolution Neural Network (CNN) YOLO series untuk beberapa variannya agar memperoleh inference ketiga model tersebut terhadap dataset mixed-traffic. Pada evaluasi pretrained skenario siang hari, diperoleh performa deteksi tebaik oleh YOLOv7-W6 dengan accuracy 0,21. Dari segi FPS, model YOLOv8s unggul dengan nilai 38,9. Untuk skenario malam hari, diperoleh performa deteksi terbaik dimiliki YOLOv7-W6 dan YOLOv8l dengan accuracy 0,18. Dari sisi FPS, model YOLOv8n mencapai nilai tertinggi sebesar 31,2. Pada evaluasi setelah trained dengan skenario siang hari, diperoleh performa deteksi terbaik model YOLOv7x, YOLOv8s, dan YOLOv8m dengan accuracy sebesar 0,95. Dari segi FPS, YOLOv7-W6 unggul dengan nilai 32,9. Untuk skenario malam hari, accuracy terbaik diperoleh model YOLOv7-E6 dan YOLOv8m sebesar 0,85. Dari sisi FPS, YOLOv7-W6 unggul dengan nilai 39,5 dari skenario siang hari.

Keyword: Mixed-traffic, Object Detection, YOLOv6, v7, v8

Mixed-traffic is a type or model of traffic that involves different types of vehicles sharing the same road infrastructure without any physical barriers or dividers. This kind of traffic environment is commonly found in urban areas in developing countries such as China, India, and Indonesia. In a mixed traffic environment, there are many different types of objects that need to be further identified. The complexity of this environment requires accurate object detection. However, the object detection used has not been optimized. Visibility in mixed traffic conditions is often low, objects in the environment are moving in the same or opposite direction, which further increases the complexity of the traffic scenario on a given road. The purpose of this research is to evaluate the performance between pretrained and trained models based on Convolution Neural Network (CNN) YOLO series for several variants in order to obtain the inference of the three models on mixed-traffic datasets. In the pretrained evaluation of the daytime scenario, the best detection performance was obtained by YOLOv7-W6 with an accuracy of 0,21. In terms of FPS, the YOLOv8s model excels with a value of 38,9. For the night scenario, the best detection performance is obtained by YOLOv7-W6 and YOLOv8l with an accuracy of 0,18. In terms of FPS, the YOLOv8 model achieved the highest value of 31,2. In the evaluation after being trained with a daytime scenario, the best detection performance was obtained for the YOLOv7x, YOLOv8s, and YOLOv8m models with an accuracy of 0,95. In terms of FPS, YOLOv7-W6 excels with a value of 32,9. For the night scenario, the best accuracy was obtained by the YOLOv7-E6 and YOLOv8m models at 0,85. In terms of FPS, YOLOv7-W6 excels with a value of 39,5 from the daytime scenario. Keyword: Mixed-traffic, Object Detection, YOLOv6, v7, v8