Robust visual voice activity detection using local variance histogram in vehicular environments

In this paper, a Vision based Voice Activity Detection (VVAD) algorithm is proposed using Local Variance Histogram (LVH). In conventional VVAD algorithm, the motion measure such as optical flow and intensity histogram are widely used. However, this approach is unstable under varying illumination and global motion changes which frequently occur in moving vehicular environment. To mitigate this problem, an appropriate framework based on LVH feature is developed. Comparison with two other conventional visual voice activity detectors shows the proposed method to be consistently more accurate and yields a substantial improvement in terms of detection probability and false alarm rate.