A new approach for hand-waving detection in crowds

In this paper, we propose a new approach to detect hand-waving motion in crowds. Different from previous approaches which are often based on segmentation and motion detection, our method can be seen as a complexity reduction process from the problem of 3D motion detection to 2D object detection. Through arranging the same row of per video frame along the time sequence, we obtain 2D images composed of traces of moving object, and in these images hand-waving motion are represented as 2D patterns which have strong periodicity. Considering the similarity between these patterns, we first use Chamfer Distance Matching to locate the patterns´ position roughly. Then we convert 2D patterns to 1D signals and apply Fourier Transformation to these signals. Finally, we identify a region as a hand-waving one, if its intensity spectrum of Fourier Transformation of 1D signals has a peak response between specified locations, and the value of the peak is greater than a threshold. Experiments on two self-built hand-waving datasets demonstrate that the proposed method can detect hand-waving motion effectively.