Action classification on the Berkeley Multimodal Human Action Dataset (MHAD)

Hu Ng*, Timothy Tzen Vun Yap, Hau Lee Tong, Chiung Ching Ho, Lay Kun Tan, Wan Xin Eng, Seng Kuan Yap, Jia Hao Soh

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

3 Citations (Scopus)


The objective of this study is to classify multimodal human actions of the Berkeley Multimodal Human Action Database (MHAD). Actions from accelerometer and motion capture modals are utilized in this study. Features extracted include statistical measures such as minimum, maximum, mean, median, standard deviation, kurtosis and skewness. Feature extraction level fusion is applied to form a feature vector comprising two modalities. Feature selection is implemented using Particle Swarm Optimization (PSO) Tabu and Ranker. Classification is performed with Support Vector Machine (SVM) Random Forest (RF) k-Nearest Neighbour (k-NN) and Best First Tree (BFT). The classification model that gave the highest accuracy is support vector machine with radial basis function kernel with a Correct Classification Rate (CCR) of 97.6 % for the Accelerometer modal (Acc) 99.8% for the Motion capture system modal (Mocap) and 99.8% for the Fusion Modal (FusioMA). In the feature selection process, ranker selected every single extracted feature (162 features for Acc and 1161 features for Mocap and 1323 features for FusioMA) and produced an average CCR of 97.4%. Comparing with PSO (68 features for Acc, 350 features for Mocap and 412 features for FusioMA) it produced an average CCR of 97.1% and Tabu (54 features for Acc, 199 features for Mocap and 323 features for FusionMA) produced an average CCR of 97.2%. Although, Ranker gave the best result, the difference in the average CCR is not significant. Thus, PSO and Tabu may be more suitable in this case as the reduced feature set can result in computational speedup and reduced complexity. The extracted statistical features are able to produce high accuracy in classification of multimodal human actions. The feature extraction level fusion to combine the two modalities performs better than single modality in the classification.

Original languageEnglish
Pages (from-to)520-526
Number of pages7
JournalJournal of Engineering and Applied Sciences
Issue number3
Publication statusPublished - 2017


  • Feature extraction
  • Feature selection
  • Human action classification
  • Machine learning
  • Multimodal

ASJC Scopus subject areas

  • General Engineering


Dive into the research topics of 'Action classification on the Berkeley Multimodal Human Action Dataset (MHAD)'. Together they form a unique fingerprint.

Cite this