Penerapan Extended Kalman Filter Untuk Sensor Fusion Kamera-LiDAR Dalam Sistem Persepsi Visual Pada Kendaraan Otonom

Alfizar, Rezi Rafidan (2025) Penerapan Extended Kalman Filter Untuk Sensor Fusion Kamera-LiDAR Dalam Sistem Persepsi Visual Pada Kendaraan Otonom. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009211024-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Kemajuan teknologi kendaraan otonom menuntut sistem persepsi yang akurat dan andal untuk pelacakan objek di lingkungan dinamis. Penelitian ini mengembangkan sistem sensor fusion yang mengintegrasikan data dari kamera dan LiDAR menggunakan Extended Kalman Filter (EKF). Sistem mencakup modul asosiasi data berbasis algoritma Hungarian, skema predictor-corrector dengan vektor status berdimensi empat, serta manajemen pelacakan yang efisien. Pengujian dilakukan pada simulator CARLA versi 0.9.13 dengan sepuluh skenario berbeda. Kamera mencapai nilai Intersection over Union (IoU) rata-rata sebesar 0,917, sedangkan sistem berbasis LiDAR saja mencatat Mean Absolute Error (MAE) sebesar 1,086 meter dengan kesalahan relatif 9,06%. Sistem sensor fusion menghasilkan MAE sebesar 1,260 meter, yang menunjukkan penurunan akurasi sebesar 16,0%. Meskipun demikian, konsistensi pelacakan tetap terjaga, sebagaimana ditunjukkan oleh standar deviasi yang serupa (0,552 meter vs. 0,485 meter). Keunggulan utama sistem ini terletak pada kemampuannya dalam menjaga identitas objek secara konsisten, mempertahankan konsistensi temporal, serta ketahanan terhadap kegagalan sensor individu. Sistem ini dapat beroperasi secara real-time dengan rata-rata kecepatan 24,6 frame per second (FPS) dan waktu pemrosesan 44,2 milidetik per frame. Penggunaan CPU tercatat hingga 25,6% secara paralel, dengan komponen visualisasi bird’s eye view menjadi beban terberat (52,9%). Hasil ini menunjukkan bahwa pendekatan sensor fusion berbasis EKF merupakan solusi yang efektif dan andal untuk pelacakan objek dalam sistem kendaraan otonom.
==========================================================================================================================================
The advancement of The advancement of autonomous vehicle technology demands accurate and reliable perception systems for object tracking in dynamic environments. This study develops a sensor fusion system that integrates data from a camera and LiDAR using an Extended Kalman Filter (EKF). The system includes a data association module based on the Hungarian algorithm, a predictor-corrector scheme with a four-dimensional state vector, and efficient tracking management. Experiments were conducted using the CARLA simulator version 0.9.13 across ten different scenarios. The camera achieved an average Intersection over Union (IoU) of 0.917, while the LiDAR-only system recorded a Mean Absolute Error (MAE) of 1.086 meters with a relative error of 9.06%. The sensor fusion system produced an MAE of 1.260 meters, indicating a 16.0% reduction in accuracy. Nevertheless, the tracking consistency remained comparable, as indicated by similar standard deviations (0.552 meters vs. 0.485 meters). The main advantage of the proposed system lies in its ability to maintain object identity over time, ensure temporal consistency, and provide robustness against individual sensor failures. The system operates in real time with an average processing speed of 24.6 frames per second (FPS) and a per-frame processing time of 44.2 milliseconds. CPU usage reached up to 25.6% in parallel execution, with the bird’s eye view visualization component contributing the highest computational load (52.9%). These results demonstrate that the EKF-based sensor fusion approach is an effective and reliable solution for object tracking in autonomous vehicle systems.vehicle technology demands accurate and reliable perception systems for object tracking in dynamic environments. This study develops a sensor fusion system that integrates data from a camera and LiDAR using an Extended Kalman Filter (EKF). The system includes a data association module based on the Hungarian algorithm, a predictor-corrector scheme with a four-dimensional state vector, and efficient tracking management. Experiments were conducted using the CARLA simulator version 0.9.13 across ten different scenarios. The camera achieved an average Intersection over Union (IoU) of 0.917, while the LiDAR-only system recorded a Mean Absolute Error (MAE) of 1.086 meters with a relative error of 9.06%. The sensor fusion system produced an MAE of 1.260 meters, indicating a 16.0% reduction in accuracy. Nevertheless, the tracking consistency remained comparable, as indicated by similar standard deviations (0.552 meters vs. 0.485 meters). The main advantage of the proposed system lies in its ability to maintain object identity over time, ensure temporal consistency, and provide robustness against individual sensor failures. The system operates in real time with an average processing speed of 24.6 frames per second (FPS) and a per-frame processing time of 44.2 milliseconds. CPU usage reached up to 25.6% in parallel execution, with the bird’s eye view visualization component contributing the highest computational load (52.9%). These results demonstrate that the EKF-based sensor fusion approach is an effective and reliable solution for object tracking in autonomous vehicle systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Extended Kalman Filter, Kamera, Kendaraan Otonom, LiDAR, Sensor fusion, Autonomous Vehicle, Camera, Extended Kalman Filter, LiDAR, Sensor fusion
Subjects:	Q Science > QA Mathematics > QA402.3 Kalman filtering. T Technology > TA Engineering (General). Civil engineering (General) > TA593.35 Instruments, cameras, etc. T Technology > TL Motor vehicles. Aeronautics. Astronautics > TL152.8 Vehicles, Remotely piloted. Autonomous vehicles.
Divisions:	Faculty of Industrial Technology > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Rezi Rafidan Alfizar
Date Deposited:	04 Aug 2025 04:14
Last Modified:	04 Aug 2025 04:16
URI:	http://repository.its.ac.id/id/eprint/126186

Actions (login required)

View Item