Improvement of vehicles accident detection using object tracking with U-Net

Abstract

Over the past decade, researchers have suggested many methods to find anomalies. However, none of the studies has applied frame reconstruction with Object Tracking (OT) to detect anomalies. Therefore, this study focuses on road accident detection using a combination of OT and U-Net associated with variants such as skip, skip residual and attention connections. The U-Net algorithm is developed for reconstructing the images using the UFC-Crime dataset. Furthermore, YOLOV4 and DeepSort are used for object detection and tracking within the frames. Finally, the Mahalanobis distance and the reconstruction error (RCE) are determined using a Kalman filter and the U-Net model.