| Summary: | Frequent object counting at a specific location (FOC@Loc) is becoming a newly emerging but highly demanded task since the evolution of human activities can provide crucial statistics for social and economic development. Due to the unique requirement of both high temporal frequency for frequent observations and high spatial resolution for object counting, this article aims to propose a novel framework for FOC@Loc to take advantage of both high definitions of high-resolution (HR) image and continuous low-cost low-resolution (LR) image at the same location. To compensate for the low ground sample distance (GSD) in LR images, some prior knowledge about the fixed location is extracted from HR images, including (1) the short-term spatial consistency: provide exact feature-wise and pixel-wise guidance from HR image to the LR image on the same date, learning to predict vehicle area for each LR image; (2) the long-term location consistency: provide the prior parking density of the study location from the sparse HR image sequence, turning the vehicle area into the counting number of each LR image. The final results indicate that this new method obtains highly consistent counting results with the manual annotations, proofing that the HR image guidance information can promote the utilization of LR images for object counting and further analysis. The proposed dataset and methodology have the potential to boost the applications of extensive and inexpensive LR satellite images.
|
|---|