摘要
1.Introduction Computer vision algorithms have attained significant accuracy in the past decade,among which arguably the most important one is deep neural networks.Unmanned aerial vehicles,commonly called drones,equipped with cameras,offer a convenient,efficient,and cost-effective way of collecting a large set of images.Combining drones and computer vision algorithms can automate the monitoring and surveying of infrastructure systems,for example,car detection(Maria et al.,2016),pedestrian and bicycle volume data collection(Kim,2020),and road degradation survey(Leonardi et al.,2018).However,the existing research has been largely driven by two independent streams of expertise:computer vision and drone scheduling.Computer scientists strive to design more accurate computer vision algorithms without much consideration of how the images are collected,whereas operations researchers endeavor to design drone routing algorithms to collect a given set of images in the most efficient manner.We suggest that the planning of images to collect(number and locations of images,amongst others)and the design of—more often than not,the choice of—computer vision algorithms should be determined holistically instead of independently.Section 2 presents an example to show the number of images to collect depends on the accuracy of the computer vision algorithms.Section 3 lays out the roadmap for future research direction.
基金
substantially supported by the funding for Projects of Strategic Importance of The Hong Kong Poly-technic University(Project Code:1-ZE2D).
