Repetitive motion detection for human behavior understanding from video images

This paper aims to develop a technique for repetitive motion detection which is necessary for human behavior analysis particularly in children with autism spectrum disorders. Images from video sequences are mainly investigated. The technique uses image self-similarity measure, which is less sensitive to view changes, noise, and stable to low resolution images, as input data to multilayer perceptron neural network. Outputs of the network are composed of two classes, which are repetitive and non-repetitive motions. The classifier uses training data from a single person. The model is created by 10 fold cross validation. Trained network is tested with different data sets from seven normal subjects. The classification results show that the proposed technique provides an average accuracy of 0.9115 and can be used in real-time manner. In addition, the trained classifier is robust to images taken from different view.