Talking Detection in Collaborative Learning Environments

We study the problem of detecting talking activities in collaborative learning videos. Our approach uses head detection and projections log-magnitude optical flow vectors to reduce a simple classification small projection images without need for training complex, 3-D activity systems. The are then easily classified using majority vote standard classifiers. For detection, our proposed is shown significantly outperform single have an overall accuracy 59% compared 42% Temporal Segment Network (TSN) 45% Convolutional 3D (C3D). In addition, method able detect multiple instances from speakers, while also speakers themselves.