VISION BASED ACCIDENT IDENTIFICATION IN TRAFFIC VIDEOS USING DEEP LEARNING

Traffic collisions have emerged as a prominent factor contributing to a significant rise in injuryrelated fatalities and injuries, making them a pressing concern for public safety. Statistics show that an estimated 1.4 million people die in traffic accidents every year and around fifty million are injured. With the increasing population, ensuring road safety has become one of the critical challenges for city administration. In response to this challenge, the field of computer vision has gained prominence, focusing on the development of digital systems capable of processing, analyzing, and interpreting visual data like humans. Recently, a few studies have applied machine learning techniques for accident detection in traffic videos, offering automated solutions for city monitoring. However, these existing models are limited to binary classification, that is, they only determine whether the accident has occurred or not. This research takes a step further by proposing a deep learning-based accident classification model that not only identifies vehicle accidents but also categorizes them based on the type of collision. We employ the transfer learning technique and evaluate five pre-trained models, including DenseNet121, InceptionNetV3, ResNet50, VGG16, and Xception. Our findings demonstrate that VGG16 stands out for the vehicle accident classification task, achieving an impressive accuracy of 96.76%. This research offers a significant contribution to enhance road safety by advancing the capabilities of accident detection and classification systems. The identification of collision types presented in this research empowers safety authorities and organizations to conduct more effective safety analyses serving dual purposes. Firstly, it ensures that emergency response teams and medical personnel are adequately prepared and secondly, it allows insurance companies to make fair determinations of liability and coverage.