Suaeb, Januar Kailani (2026) Pengembangan Pipeline Terintegrasi Berbasis Kamera Dashcam Untuk Prediksi Near-Miss Menggunakan Metode Stokastik Surrogate Safety Measures. Masters thesis, Institut Teknologi Sepuluh Nopember.
|
Text
6025242009-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy |
Abstract
Prediksi nearmiss pada lalu lintas dari kamera monokuler masih menjadi tantangan tersendiri karena penilaian bahaya bersifat subjektif, kondisi jalan beragam, serta pergerakan dari kendaraan atau objek di jalan harus diukur dari kendaraan yang sama-sama ikut bergerak. Pendekatan deterministik klasik membuat logika penentuan nearmiss menjadi transparan, akan tetapi masih rentan terhadap konstanta-konstanta statis pada masing-masing SSM seperti TTC, DRAC, PET, dan MinDis dalam penentuan apakah suatu kondisi dikategorikan nearmiss atau tidak, sedangkan pendekatan dengan memanfaatkan machine learning menuntut anotasi dan komputasi yang besar. Penelitian ini membangun sebuah alur pemrosesan menyeluruh yang terdiri atas empat tahap terpisah dan dapat diukur sendiri, yaitu transformasi gambar bird’s eye view, Ego Vehicle Speed Estimation, Detected Object Motion Estimation, serta modul prediksi nearmiss. Kontribusi utama pada penelitian ini berupa modul prediksi Stochastic Monte Carlo Surrogate Safety Measure (MC-SSM) yang menghasilkan keluaran berupa probabilitas nearmiss yang kontinu. Geometri bird’s eye view diperoleh tanpa masking segmentasi melalui heatmap sudut jalan, dengan nilai corner error 91.7. Ego-speed Estimation pada penelitian ini menggunakan Temporal Recurrent Network yang memanfaatkan perhitungan Optical Flow dengan nilai Mean Absolute Error (MAE) yang didapatkan adalah 8.77 km/h. Untuk Estimasi gerak objek yang robust berbasis Theil Sen Regression berhasil menurunkan eror menjadi hanya 7.3% pada Dataset NEXAR. Sedangkan untuk prediksi nearmiss pada NEXAR dataset mendapatkan nilai Mean Average Precision (mAP) 0.651 pada seluruh seribu lima ratus klip. Penelitian ini menunjukkan bahwa peningkatan kualitas dari estimasi Object Motion adalah langkah yang sangat berpengaruh terhadap nilai kepresisian dari prediksi nearmiss. Kontribusi utama pada penelitian ini adalah framework modular, tidak terikat pada dataset nearmiss, dan transaparansi algoritma.
===============================================================================================================================
Nearmiss prediction in traffic from monocular cameras remains a challenge due to subjective hazard assessment, varying road conditions, and the movement of vehicles or objects on the road must be measured from other moving vehicles. The classical deterministic approach makes the logic of nearmiss determination transparent, but is still vulnerable to static constants in each SSM, such as TTC, DRAC, PET, and MinDis, in determining whether a condition is categorized as nearmiss or not, while the approach utilizing machine learning requires large annotations and computations. This study builds a comprehensive processing flow consisting of four separate and independently scalable stages, namely bird's eye view image transformation, Ego Vehicle Speed Estimation, Detected Object Motion Estimation, and a near-miss prediction module. The main contribution to this study is the Stochastic Monte Carlo Surrogate Safety Measure (MC-SSM) prediction module that produces output in the form of a continuous nearmiss probability. The bird's eye view geometry is obtained without segmentation masking through a road corner heatmap, with a corner error value of 91.7. Ego-speed Estimation in this study uses a Temporal Recurrent Network that utilizes Optical Flow calculations, yielding a Mean Absolute Error (MAE) of 8.77 km/h. For robust object motion estimation based on Theil-Sen Regression, it succeeded in reducing the error to only 7.3% on the NEXAR Dataset. Meanwhile, for nearmiss prediction on the NEXAR dataset, the Mean Average Precision (mAP) value was 0.651 on all one thousand five hundred clips. This study shows that improving the quality of Object Motion estimation is a very influential step on the precision value of nearmiss predictions. The main contribution of this study is the modular framework, not tied to the nearmiss dataset, and the transparency of the algorithm.
| Item Type: | Thesis (Masters) |
|---|---|
| Uncontrolled Keywords: | Bird’s Eye View, Kamera Monokuler, MC-SSM, Nearmiss, Optical Flow Bird’s Eye View, MC-SSM, Monocular Camera, Near-miss, Optical Flow |
| Subjects: | T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing. |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis |
| Depositing User: | Januar Kailani Suaeb |
| Date Deposited: | 29 Jul 2026 20:41 |
| Last Modified: | 29 Jul 2026 20:41 |
| URI: | http://repository.its.ac.id/id/eprint/139537 |
Actions (login required)
![]() |
View Item |
