V3C-NET: End to End Traffic Volume Estimation

Abstract

Visual object counting has been a well-researched and solved topic over the past few years in computer vision. Previous computer vision solutions fail to provide a robust solution to changes in illumination, occlusion, varying weather conditions and camera/vehicle orientation. These problems can be solved by providing machine learning solution. We propose an end to end method for counting of vehicular volume. We aim to replace the manual collection of the count carried out by civil engineers and surveyors. This data is important in identifying frequently used routes, roads that need more lanes, length of the lanes/roads and suitable traffic control methods. Our method uses 3D Convolutional Neural network which learns spatial-temporal features of vehicular volume and count them as the vehicle leaves the frame. Our method is able to learn to count vehicles without the use of any extra pre-processing, or applying any object tracking methods, and counting them when they pass through a line of interest or region of interest. Our method is robust to occlusions, camera/vehicle orientation, lighting condition and various weather conditions.