Efficient Vehicle Counting Based on Time-Spatial Images by Neural Networks

A highly efficient vehicle counting approach based on timespatial images with deep learning is proposed in this paper. Most vehicle counting solutions are based on frame-by frame object detection and tracking to calculate the number of cars that cross a counting line. However, these approaches incur a great deal of redundancy because they track vehicles in a large area though it matters only when vehicles cross the counting line. In this work, we use time-spatial images to focus only on the information happening along the counting lines, instead of whole images, to reduce redundancy. Due to the nature of time-spatial images, vehicle counting can be achieved by object detection in such images without frame-by-frame tracking. We propose Foreground Favorable Model to conquer occlusion, congestion, and lighting change problems and Cross-Image Object Linking to conquer the distortion problem of nearly static vehicles. We also present an automatic time-spatial image dataset generation flow and the first time-spatial image dataset, called DRIVE-TSI, for vehicle counting tasks. Our vehicle counting accuracy beats state-of-the-art solutions in accuracy and is proved to be much more efficient because it only focuses on a small number of pixels. Our model achieves a 97.95% counting accuracy at 2.91 ms per frame in day time urban scenarios.