2018/11/29 by Yaadhav Raaj, Haroon Idrees, Raaj, Yaadhav +5 · 1 citation
Computer Science · #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #Video Surveillance and Tracking Methods
paper · pdf · doi:10.48550/arxiv.1811.11975
openalex publication_date 2018/11/29 · openalex created_date 2022/08/01 · openalex updated_date 2026/07/28
We present an online approach to efficiently and simultaneously detect and\ntrack the 2D pose of multiple people in a video sequence. We build upon Part\nAffinity Field (PAF) representation designed for static images, and propose an\narchitecture that can encode and predict Spatio-Temporal Affinity Fields (STAF)\nacross a video sequence. In particular, we propose a novel temporal topology\ncross-linked across limbs which can consistently handle body motions of a wide\nrange of magnitudes. Additionally, we make the overall approach recurrent in\nnature, where the network ingests STAF heatmaps from previous frames and\nestimates those for the current frame. Our approach uses only online inference\nand tracking, and is currently the fastest and the most accurate bottom-up\napproach that is runtime invariant to the number of people in the scene and\naccuracy invariant to input frame rate of camera. Running at \∼30 fps on a\nsingle GPU at single scale, it achieves highly competitive results on the\nPoseTrack benchmarks.\n