Drone-based Object Counting by Spatially Regularized Convolutional Neural Networks
Innovations Drone-based Object Counting by Spatially Regularized Convolutional Neural Networks
  • Technology Introduction
Drone-based object counting is vital due to the prevalence of drones. We propose Layout Proposal Networks (LPNs) to simultaneously count and localize target objects (e.g., cars) in drone-view videos. The method can be extended to other valuable objects such as cows, tanks, etc. We leverage the spatial layout cues (e.g., cars often park regularly) to augment the network design. We also present a new large-scale dataset (CARPK) that contains nearly 90K cars captured from different parking lots.
 
  • Scientific Innovation
To our knowledge, this is the first work that leverages spatial layout cues for drone-view object region proposal. We improve the average recall of the state-of-the-art region proposal methods (i.e., 59.9% to 62.5%) on a public PUCPR dataset. We contributed the large-scale dataset (CARPK) containing more than 90K cars, the first drone-view dataset. Moreover, it out performs state-of-the-art object detection methods such as Faster RCNN, YOLO, etc.
 
  • Institutes
National Taiwan University
Winston Hsu / whsu@ntu.edu.tw
 

 
本網站使用您的Cookie於優化網站及您的瀏覽經驗。繼續瀏覽網站即表示您同意本公司隱私權政策,您可至隱私權政策了解詳細資訊。