2017/04/09 by Rodney LaLonde, Dong Zhang, LaLonde, Rodney +3 · 1 citation
Computer Science · #Advanced Neural Network Applications #Artificial intelligence #Background subtraction #Centroid #Computer Vision and Pattern Recognition (cs.CV) #Computer science #Computer vision #Convolutional neural network #FOS: Computer and information sciences #Frame (networking) #Human Pose and Action Recognition #Motion (physics) #Object (grammar) #Object detection #Pattern recognition (psychology) #Pixel #Set (abstract data type) #Video Surveillance and Tracking Methods #Visual Attention and Saliency Detection #cs.CV
paper · pdf · doi:10.48550/arxiv.1704.02694
published in arXiv (Cornell University) (Cornell University) · Main paper is 8 pages. Supplemental section contains a walk-through of our method (using a qualitative example) and qualitative results for WPAFB 2009 dataset
openalex publication_date 2017/04/09 · arxiv created 2017/12/04 · arxiv updated 2017/12/06 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Object detection in wide area motion imagery (WAMI) has drawn the attention\nof the computer vision research community for a number of years. WAMI proposes\na number of unique challenges including extremely small object sizes, both\nsparse and densely-packed objects, and extremely large search spaces (large\nvideo frames). Nearly all state-of-the-art methods in WAMI object detection\nreport that appearance-based classifiers fail in this challenging data and\ninstead rely almost entirely on motion information in the form of background\nsubtraction or frame-differencing. In this work, we experimentally verify the\nfailure of appearance-based classifiers in WAMI, such as Faster R-CNN and a\nheatmap-based fully convolutional neural network (CNN), and propose a novel\ntwo-stage spatio-temporal CNN which effectively and efficiently combines both\nappearance and motion information to significantly surpass the state-of-the-art\nin WAMI object detection. To reduce the large search space, the first stage\n(ClusterNet) takes in a set of extremely large video frames, combines the\nmotion and appearance information within the convolutional architecture, and\nproposes regions of objects of interest (ROOBI). These ROOBI can contain from\none to clusters of several hundred objects due to the large video frame size\nand varying object density in WAMI. The second stage (FoveaNet) then estimates\nthe centroid location of all objects in that given ROOBI simultaneously via\nheatmap estimation. The proposed method exceeds state-of-the-art results on the\nWPAFB 2009 dataset by 5-16% for moving objects and nearly 50% for stopped\nobjects, as well as being the first proposed method in wide area motion imagery\nto detect completely stationary objects.\n