Implementasi I3D untuk Deteksi Anomali pada Video Dashcam

Saputra, Dana (2025) Implementasi I3D untuk Deteksi Anomali pada Video Dashcam. Project Report. [s.n.], [s.l.]. (Unpublished)

Text 5025221003-Project_Report.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Kecelakaan lalu lintas merupakan salah satu penyebab utama kematian di dunia serta menimbulkan kerugian ekonomi yang signifikan setiap tahunnya. Seiring dengan meningkatnya jumlah kendaraan dan penggunaan sistem pemantauan berbasis kamera seperti dashcam video, dibutuhkan sistem deteksi otomatis yang mampu mengenali kejadian anomali atau kecelakaan secara cepat dan akurat. Penelitian ini mengimplementasikan metode Inflated 3D ConvNet (I3D) untuk mendeteksi anomali pada video dashcam. Model I3D yang digunakan merupakan model pretrained pada dataset Kinetics dan UCF-Crime, dengan dua flow ekstraksi fitur, yaitu RGB stream dan Optical Flow stream. Masing-masing flow menghasilkan vektor fitur berdimensi 1024, sehingga total fitur gabungan berukuran 2048. Fitur tersebut kemudian dilatih menggunakan fully connected learner berarsitektur tiga layer untuk mengklasifikasikan segmen video ke dalam kategori normal atau anomali. Eksperimen dilakukan menggunakan 507 video anomali dan 479 video normal dari dataset dashcam yang telah dianotasi secara manual. Evaluasi pada level frame menunjukkan nilai AUC sebesar 0,77, recall 0,85, dan precision 0,22 pada fixed threshold. Hasil ini menunjukkan bahwa model memiliki sensitivitas tinggi terhadap kejadian anomali, namun masih menghasilkan tingkat false positive yang relatif tinggi. Sebagai output akhir, ditampilkan visualisasi skor anomali per frame dalam bentuk grafik dan video overlay menggunakan OpenCV, dengan penandaan bingkai merah pada frame yang terdeteksi sebagai anomali. Penelitian ini menunjukkan bahwa pendekatan berbasis I3D dengan two-stream feature extraction efektif untuk mendeteksi kejadian anomali pada video dashcam, serta membuka peluang pengembangan lebih lanjut menuju sistem deteksi kecelakaan otomatis di masa mendatang.
====================================================================================================================================
Traffic accidents are one of the leading causes of death worldwide and result in significant economic losses each year. With the increasing number of vehicles and the widespread use of camera-based monitoring systems such as dashcam videos, there is a growing need for an automated detection system capable of identifying anomalous or accident events quickly and accurately. This study implements the Inflated 3D ConvNet (I3D) method to detect anomalies in dashcam videos. The I3D model used is pretrained on the Kinetics and UCF-Crime datasets, employing two feature extraction flows: the RGB stream and the Optical Flow stream. Each flow produces a 1024-dimensional feature vector, resulting in a combined 2048-dimensional feature representation. These features are then trained using a fully connected learner with a three-layer architecture to classify video segments into normal or anomalous categories. Experiments were conducted using 507 anomalous videos and 479 normal videos from a manually annotated dashcam dataset. Frame-level evaluation achieved an AUC of 0.77, recall of 0.85, and precision of 0.22 at a fixed threshold. These results indicate that the model exhibits high sensitivity in detecting anomalous events but still produces a relatively high false positive rate. As the final output, anomaly scores are visualized per frame in the form of both a graph and a video overlay using OpenCV, with red bounding boxes marking frames detected as anomalous. This study demonstrates that the I3D-based two-stream feature extraction approach is effective for detecting anomalous events in dashcam videos and provides a promising foundation for the further development of automated accident detection systems in the future.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Advanced Driver Assistance Systems (ADAS), Anomaly Detection, I3D Advanced Driver Assistance Systems (ADAS), Anomaly Detection, I3D
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Dana Saputra
Date Deposited:	13 Nov 2025 01:19
Last Modified:	13 Nov 2025 01:19
URI:	http://repository.its.ac.id/id/eprint/128784

Actions (login required)

View Item