Cal Poly
Leveraging Machine-Learning Algorithms in Two Car Crash Detection Systems on a Custom Dataset
Abstract
dc:description.abstractTraffic accidents pose a significant threat to public safety, causing millions of deaths and injuries worldwide each year. While efforts to reduce accidents have seen limited progress in recent years, improving emergency response times through automated detection systems is a promising avenue for saving lives. This thesis describes the development of machine learning-based traffic accident detection systems, exploring both video classification and image detection models. The models are trained on a new dataset deemed the Cal Poly Traffic Accident Dataset, an extension of the existing Car Accident Detection and Prediction (CADP) dataset with a precise collision annotations. Two systems were developed and evaluated. The first is centered around the R(2+1)D video classification model, in which the system processes temporal information in isolated regions of videos. And the second, is the YOLOv8-nano image detection model, which detects accidents on a frame-by-frame basis without temporal context. Experimental results revealed that despite lacking temporal awareness, the YOLO-based model outperformed the system with the R(2+1)D model. This suggests that ensuring strong spatial feature extraction is crucial before incorporating temporal processing in crash recognition systems.
Degree
thesis:*- Name thesis:degree_name
- MS in Electrical Engineering
- Discipline thesis:degree_discipline
- Electrical Engineering
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sandvik, Addison Jacob
- Contributors dc:contributor
-
- Jane Zhang
- Electrical Engineering
- College of Engineering
Subjects
dc:subject × 7Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.calpoly.edu/theses/3098