Research on Vehicle Detection and Direction Determination based on Deep Learning

Abstract

With the increase of vehicle ownership in China, the number of auto insurance cases is also increasing. The detection and direction determination of vehicles involved in auto insurance cases have important applications in the field of intelligent loss assessment. In this paper, a model of vehicle detection and direction determination based on ResNet-101+FPN backbone network and RetinaNet is built by using convolutional neural network in deep learning. Then, the model is trained and tested on the labelled data set. The model has a relatively high accuracy of prediction, in which the accuracy of vehicle detection reaches 98.7%, and the accuracy of the five directions determination of frontal, lateral-frontal, lateral, lateral-back and back reaches 97.2%.