A novel hybrid deep learning model for earlier accident prediction using computer vision on surveillance camera
Supplementary material
Other Title
Authors
Mojumder, Atanu
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2024
Supervisors
Keivanmarz, Ali
Pashna, Mohsen
Pashna, Mohsen
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
traffic accidents
traffic safety
computer vision
object recognition
deep learning
real-time data processing
neural networks
traffic safety
computer vision
object recognition
deep learning
real-time data processing
neural networks
ANZSRC Field of Research Code (2020)
Citation
Mojumder, A. (2024). A novel hybrid deep learning model for earlier accident prediction using computer vision on surveillance camera (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology.
https://hdl.handle.net/10652/6747
Abstract
Frequent incidence of road accidents worldwide drives continuous efforts toward accident prediction to ensure road safety. Road accidents pose a significant global issue that causes millions of causalities with huge economic losses. Human errors, unsafe roads, and risky weather conditions usually cause those accidents. Early accident detection is a significant factor in preventing road accidents by enabling real-time alerts and faster response and reducing the causality from road accidents. Using advanced computer vision and object recognition models, early accident detection systems provide practical solutions for preventing road accidents using real-time alerts and faster responses.
In this research, we’ll present a hybrid model that can detect accidents earlier by focusing on improving accuracy and prediction time. We collected videos from the CADP dataset and used a custom dataset for forecasting, and we pre-processed the data by labeling, modifying, and converting it into images. The study reviews computer vision, machine learning methods, and technologies used in earlier traffic accident prediction by proposing a hybrid model using multiple versions of the YOLO model, like YOLOv8, YOLOv10, YOLOv10, YOLOv11, and Faster R-CNN. We trained all those models on fifty custom datasets multiple times with different parameters for better detection accu racy and output.
In the result analysis, the individual models like YOLOv8, YOLOv10, YOLOv11, and Faster R-CNN, achieved an accuracy of 57.14%, 60%, and 42.86% for both YOLOv11 and Faster R-CNN, respectively. In individual comparison, YOLOv10 has the best maximum detection time at 2.228 seconds, and YOLOv11 has the best average accident detection time at 1.125 seconds. From comparison, the hybrid model has the best performance, with the best average detection time of 2.067 seconds and the best maximum detection time of 3.567 seconds. The proposed Hybrid model contains three trained models, which consist of YOLOv8, YOLOv10, and Faster R-CNN, with an accuracy of accident forecasting 88.57% accuracy, where this model predicts thirty-one incident videos out of thirty-five incidents in video data
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved