Density Map Augmentation-Based Point-to-Point Vehicle Counting and Localization in Remote Sensing Imagery with Limited Resolution

Abstract

Monitoring traffic flow is of great significance to contemporary urban management and intelligent transportation construction. Among them, satellite remote sensing images are the most easily accessible and cost-effective remote sensing image data sources. However, most high-resolution satellite remote sensing images tend to have only sub-meter (0.5 m–1 m) spatial resolution, which makes vehicle extraction from such images very difficult to be carried out, and there are only a few related research studies. If vehicle counting and localization can be done under such limited-resolution remote sensing images, the application potential and study’s value of such images can be greatly explored, and large-scale multi-temporal urban traffic condition analysis can be conducted with low acquisition costs. After the vehicle counting work with limited-resolution remote sensing images is proven to be feasible, this paper will improve the granularity of vehicle extraction tasks and conduct research studies on vehicle counting and localization tasks under this resolution. Combining the existing research studies results on density map counting, this paper designs a density map enhancement module that can be used to enhance the vehicle localization model and proposes an improved model based on the classical point-based keypoint object localization model P2PNet. The effectiveness of the proposed point-based keypoint localization model based on the density map enhancement strategy is verified on the RSVC dataset for the vehicle localization task of limited-resolution remote sensing images.