Proposal for automating vehicle gauging from road images using convolutional neural networks

Abstract

The knowledge of the vehicular flow of a road is an essential factor in road planning. Therefore, in the present study, a deep learning model based on the Faster-RCNN neural network architecture is proposed to get the vehicle capacity of a road without the need for human supervision. The experimental research methodology applied allowed testing and optimizing the candidate model with the gradient descent algorithm to locate vehicles and determine the vehicle capacity in road images. The proposed model receives as input an image represented in a matrix and outputs the image with the localized vehicles. Once the model was trained with highway images in China, its performance was evaluated with pictures of roads in Ecuador which were captured by cameras of the ECU911 available on the web. As a result of the evaluation, a functional deep learning model was obtained for automating vehicle counting with an accuracy of 95%. Even though the model was trained with images from another country, the results show the model adapts favorably to the reality of Ecuador with an optimal inference time of 0.28 seconds.