Two person interaction recognition based on a dual-coded modified metacognitive (DCMMC) extreme learning machine
Turkish Journal of Electrical Engineering and Computer Sciences, cilt.30, sa.4, ss.1621-1636, 2022 (SCI-Expanded, Scopus, TRDizin)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 30 Sayı: 4
- Basım Tarihi: 2022
- Doi Numarası: 10.55730/1300-0632.3869
- Dergi Adı: Turkish Journal of Electrical Engineering and Computer Sciences
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Applied Science & Technology Source, Compendex, Computer & Applied Sciences, INSPEC, TR DİZİN (ULAKBİM)
- Sayfa Sayıları: ss.1621-1636
- Anahtar Kelimeler: Human interaction recognition, metacognition, extreme learning machines, dual coding theory, dense trajectories
- Atatürk Üniversitesi Adresli: Hayır
Özet
Human action recognition has been an active research area for over three decades. However, state-of-the-art proposed algorithms are still far from developing error-free and fully-generalized systems to perform accurate interaction recognition. This work proposes a new method for two-person interaction recognition from videos, based on well-known cognitive theories. The main idea is to perform classification based on a theory of cognition known as dual coding theory. The theory states that human brain processes and represents two types of information to learn/classify data named analogue and symbolic codes, i.e. (verbal as analogue and visual as symbolic). To implement such a theory in a two-person interaction classification system, we exploit dense trajectories as analogue codes and a bag of words as symbolic codes which are two code types hypothesized in the theory. In addition to dual coding theory, we propose to implement a metacognitive classifier model which adds a metalevel with its own rules to perform more accurate training process. We also propose a modification in a metacognitive component to prevent cognitive interference well known as the Stroop effect. Evaluations on both datasets revealed that the method offers comparable recognition accuracy (95.6% for the SBU interaction dataset and 91.1% for the UT-interaction dataset).