Quantized feature with angular displacement for activity recognition

Abstract

We propose quantized feature with angular displacement for pose-based activity recognition. We calculate a 3D joint angle from three postural coordinates. The angular displacement should be quantized since joint angle includes errors due to system noises and similar posture. To investigate appropriate features, we propose four kinds of quantization levels; binarization, ternarization, quaternarization, and quinarization. We apply quantized features in order to improve pose-based activity recognition with the UTKinect-Action Dataset. In the experiment, we show the appropriate feature for activity recognition. As the result, the ternarized feature achieves the highest recognition rate in average. The recognition rate of trials with ternarized feature is improved 2.4% to one with no-quantized feature, and 1.8% to conventional method.