Recognizing Multi-Agent Activities from GPS Data

• DOI: 10.1609/aaai.v24i1.7739
• Publication Date: 2022-09-13T04:59:59Z
• Authors: Adam Sadilek; Henry Kautz
• Keywords: 03 medical and health sciences; 0302 clinical medicine

Recent research has shown that surprisingly rich models of human behavior can be learned from GPS (positional) data. However, most to date concentrated on modeling single individuals or aggregate statistical properties groups people. Given noisy real-world data, we---in contrast---consider the problem and recognizing activities involve multiple related playing a variety roles. Our test domain is game capture flag---an outdoor involves many distinct cooperative competitive joint activities. We model using Markov logic, relational language, learn theory jointly denoises data infers occurrences high-level activities, such as capturing player. combines constraints imposed by geometry area, motion players, rules dynamics in probabilistically logically sound fashion. show while it may impossible directly detect multi-agent activity due sensor noise malfunction, occurrence still inferred considering both its impact future behaviors people involved well events could have preceded it. compare our unified approach with three alternatives (both probabilistic nonprobabilistic) where either denoising detection are strictly separated, states players not considered, both. time window spanning entire game, although more computationally costly, significantly accurate.