3-D Tracking Research: Learning Correspondence from Static 3D Points Causes 3D Object Tracking to Emerge
Abstract
Problem: Aerial surveillance demands full attention to video by PED teams Manual, error-prone process Technical barriers including object detection, and tracking Limitations result in poor pattern detection in a surveilled region Vehicle tracks used to train LSTM autoencoder that learns normal behavior in order to identify anomalous tracks Results shown are for perfect data -- reality is not so pretty due to inadequate object detection and tracking This results in lost tracks and many track lets that are difficult to correlate Solution Work directly with DoD to improve pattern detection in aerial surveillance data patterns Work with researchers to address core technology problems of tracking of objects Impact (FY1820) Improved DoD pattern detection in aerial surveillance data Developing unsupervised 3D tracking algorithms to improve on other unsupervised methods and achieve performance similar to supervised methods
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2020
- Accession Number
- AD1110827
Entities
Organizations
- Carnegie Mellon University