Vehicle Collision Prediction from Visual Sizes of Objects via Car Dash Camera

The purpose of this research is to develop a method for vehicle collision prediction from visual sizes of objects. There are three main stages in the proposed method. The first stage is object detection and classification using You Only Look Once version 7 (YOLOv7) and object tracking using Simple Online and Realtime Tracking with a deep association metric (DeepSORT). Then, linear regression is applied with the data gathered from the first stage to predict the shape of the bounding box and the coordinates of the centroid of the bounding box of the object at various time intervals. The last stage is vehicle collision prediction along with the time interval in which the collision may occur. If the predicted bounding box of the object overlaps with the defined area in the video, the proposed method predicts that the collision will occur. From the experimental results, the proposed method can trigger collision warning and help reduce the risk of road accidents with car dash camera.