Automatic Detection of Road Accidents from Surveillance Videos Using Spatial Temporal Vision Transformer with Adaptive EfficientNetB7

Abstract

Global data elucidate that a significant number of violent deaths are caused by unexpected accidents. Automatic accident detection, particularly through video analysis and paves more attention in the past few years. It is crucial for both traffic control and Intelligent Transportation Systems (ITS) since it prevents accidents from getting worse, particularly on highways. Numerous cameras have been installed in some kinds of public and private locations for traffic monitoring, surveillance, and the tracking of unusual human behavior. Due to a lack of accident data for training, using machine learning and computer vision techniques to identify traffic accidents is a difficult challenge. In this work, a road accident detection model using advanced deep learning techniques is proposed to allow quick emergency response. The input surveillance videos are collected from appropriate sites. These collected videos are converted into frames to ensure better consistency during accident identification. The converted frames are used for road accident detection using the proposed Spatial Temporal Vision Transformer with Adaptive EfficientNetB7 (STViT-AEB7). Furthermore, the efficiency of this network is improved by optimizing the parameters using the Fitness Solutions-based Randomized Giant Trevally Optimizer (FSR-GTO). The suggested model helps to prevent the loss of life by providing immediate response and improving road safety. The outcomes achieved by the designed model are compared with traditional algorithms and approaches to ensure the model’s effectiveness.