ENSEMBLE DEEP LEARNING FRAMEWORK FOR TRAFFIC ACCIDENT DETECTION IN SMART CITIES

Abstract

Effective accident detection techniques are essential for improving safety and expediting traffic management in smart cities due to the dynamic and unpredictable nature of road traffic. In addition to providing a thorough overview of various traffic accident types, such as rear-end collisions, T-bone collisions, and frontal impact accidents, this paper provides an in-depth investigation study of popular accident detection techniques, illuminating the subtleties of other cutting-edge approaches. By combining RGB frames with optical flow data, our innovative method presents the I3D-CONVLSTM2D model architecture, a lightweight solution specifically designed for accident detection in smart city traffic surveillance systems. Our experimental study's empirical analysis highlights how effective our model design is. With a remarkable Mean Average Precision (MAP) of 87%, the I3D-CONVLSTM2D RGB + Optical-Flow (trainable) model performed better than its competitors. Our results provide more insight into the difficulties caused by data imbalances, especially when dealing with a small number of datasets, road configurations, and traffic situations. In the end, our study shows the way to an advanced vision-based accident detection system that is ready for immediate incorporation into edge IoT sensors in smart city infrastructures.