Deep learning-based traffic accident detection and severity classification in video surveillance.

Abstract

Since there has been an increase in the number of cars on the road, there is need to fast identify accident-prone areas and respond to them promptly. This Thesis aims at finding the suitability of employing state-of-the-art deep learning algorithms in identifying traffic accidents and their severity through video surveillance systems. To achieve this, using the large set of traffic videos, we learned several deep learning models which could identify whether an accident occurred based on the video or show how severe the accident was. The main model used in this research, called the Long-term Recurrent Convolutional Network (LRCN), combines two powerful techniques: The first is to distinguish between the fine details of the video at a frame level while the second is used in capturing the sequence of events in the material. Apart from the LRCN model, several other models like LSTM, 3D CNN, and MoVinet were also used for evaluation for better understanding of the proposed method’s performance and its variations with other models. The proposed LRCN model was unique by presenting a high degree of accurate identification of incidents and their severity. Processing of the videos gave it a unique chance of capturing both the spatial and temporal aspects, hence seeing patterns, which other models failed on. These findings underscore the ability of LRCN in revolutionizing how we monitor traffic accidents hence making the roads safer and with better response time to the incidents. It provides the way out to realize more sophisticated traffic management systems, which can recognize the cases of the accidents, the urgency of responses, and may subsequently reduce the number of fatalities related to the traffic incidents.