Residual Network (ResNet) Based Deep Learning Method for Detection and Classification of Accidents in Surveillance Scenes

Road accidents have become a major cause of death worldwide. Common causes of these accidents include speeding, driving while intoxicated, and unpredictable weather conditions. However, the primary cause of these accidents is the inability to detect them in a timely manner. Although human operators are being used to monitor the surveillance continuous monitoring is difficult and to errors. Recently Computer vision techniques are being deployed to detect the accidents automatically in surveillance videos. Many techniques have been developed in the field of computer vision and machine learning to detect the anomalies in the videos. However, most of the methods do not focus on detecting accidents and the state of art accuracy of the proposed models is very less. To overcome these issues a novel method is proposed using Residual Networks (ResNet) to detect the accidents in real time. Proposed ResNet architecture extracts the higher-level features from the input frames and recurrent network extract the temporal representations. Experiments have been performed on various accident detection datasets like HID12 and UCF CRIME. This study clearly distinguished between normal and abnormal behavior, demonstrating that ResNet is able to classify each accident into the appropriate group. ResNet50, ResNet101 and ResNet152 obtained 98.34% accuracy on the UCF-Crime dataset.