Model Hybrid EfficientNetV2-CaT Block untuk Pengenalan Perilaku Pengemudi Kendaraan

Kamila, Shafa (2025) Model Hybrid EfficientNetV2-CaT Block untuk Pengenalan Perilaku Pengemudi Kendaraan. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002211120-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB) | Request a copy

Abstract

Kecelakaan lalu lintas akibat perilaku pengemudi merupakan masalah global yang terus meningkat dan memerlukan solusi inovatif. Penelitian ini mengimplementasikan model hybrid berbasis deep learning, yaitu EfficientNetV2-CaT Block, untuk mengenali perilaku pengemudi kendaraan. Model ini menggabungkan keunggulan EfficientNetV2 dalam mengekstraksi fitur lokal gambar secara efisien, dengan kemampuan CaT Block yang memanfaatkan Transformer Encoder untuk memahami fitur global. Setiap CaT Block terdiri dari proses flattening fitur spasial menjadi urutan vektor, pemrosesan oleh Transformer Encoder, dan folding kembali ke bentuk spasial awal. Penelitian ini menggunakan data sekunder dari State Farm Distracted Driver Detection dan data primer yang mencakup 9 kategori perilaku pengemudi. Data diproses melalui tahapan pra-pemrosesan, meliputi resizing, normalisasi, dan konversi ke bentuk tensor, serta dilakukan augmentasi pada data latih berupa horizontal flip, rotasi, penyesuaian brightness, dan kontras. Model dilatih menggunakan data latih, dievaluasi dengan data validasi,dan diuji kinerjanya menggunakan data uji. Berbagai variasi percobaan dilakukan untuk mengkaji pengaruh penerapan augmentasi, dropout, serta skenario penempatan CaT Block dalam arsitektur hybrid. Hasil eksperimen menunjukkan bahwa model hybrid dengan dua CaT Block yang disisipkan setelah Stage 3 dan Stage 7 pada EfficientNetV2, serta penerapan augmentasi data dan dropout memberikan performa terbaik, yang ditunjukkan oleh nilai recall rata-rata sebesar 0.9711 pada seluruh kategori perilaku. Model ini memiliki potensi
besar untuk diimplementasikan sebagai sistem peringatan dini dalam kendaraan.
======================================================================================================================================
Traffic accidents resulting from driver behavior remain a pressing global issue, highlighting the need for innovative technological solutions. This study presents a hybrid deep learning model EfficientNetV2 enhanced with CaT Blocks for the classification of driver behavior based on visual input. The model combines the efficiency of EfficientNetV2 in extracting localized image features with the global contextual awareness provided by CaT Blocks, which incorporate Transformer Encoders. Each CaT Block executes a sequence of operations, including the flattening of spatial feature maps into vector sequences, the application of self-attention via a Transformer Encoder, and the reconstruction of spatial structure through a folding process. The dataset used consists of secondary data from the State Farm Distracted Driver Detection benchmark and primary data encompassing nine categories of driver behavior. The data undergoes preprocessing steps such as resizing, normalization, and tensor conversion, followed by data augmentation techniques including horizontal flipping, rotation, and adjustments to brightness and contrast. The model is trained on the augmented dataset, validated to select the optimal configuration, and tested on a separate test set. A series of experiments were conducted to evaluate the impact of data augmentation, dropout, and various CaT Block placement strategies within the hybrid architecture. Experimental results indicate that the optimal configuration incorporating two CaT Blocks after Stages 3 and 7 of EfficientNetV2, along with data augmentation and dropout, achieves an average recall score of 0.9711 across all behavior categories. These findings demonstrate the model’s strong potential for integration into real-time in-vehicle early warning systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Pengenalan Perilaku Pengemudi, EfficientNetV2, CaT Block, Dropout, Augmentasi, Driver Behavior Recognition, EfficientNetV2, CaT Block, Dropout, Augmentation
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Shafa Kamila
Date Deposited:	27 Jul 2025 02:32
Last Modified:	27 Jul 2025 02:32
URI:	http://repository.its.ac.id/id/eprint/121590

Actions (login required)

View Item