Vol-Net and V3C-Net: Towards End to End Traffic Volume Estimation

Abstract

Accurate traffic volume estimation is highly important for many important governance tasks including traffic management and city planning. In this paper, we present two deep networks for counting vehicles in traffic video streams aka traffic volume estimation. Unlike existing approaches that are based on object detection and tracking, we propose deep networks to learn spatio-temporal features that can directly estimate traffic volume. In one approach, spatial features extracted by a time-distributed convolutional neural network are temporally aggregated by recurrent neural networks to predict traffic volume. In the other approach, 3D convolutional neural network extracts the spatiotemporal features across the input frames and predicts the traffic volume. To the best of our knowledge, this is the first attempt to estimate traffic volume in an end-to-end fashion. To promote research and development for the solution of this particular problem, we contribute a challenging dataset and therefore, establish a competent baseline method for comparative analysis. The dataset is a first of its kind that provides ground truths for direct estimation of traffic volume. Our experimental results show that the proposed end-to-end methods significantly outperform the baseline approach.