Object Tracking from Audio and Video data using Linear Prediction method

Abstract

Microphone arrays and video surveillance by camera are widely used for detection and tracking of a moving speaker. In this project, object tracking was planned using multimodal fusion i.e., Audio-Visual perception. Source localisation can be done by GCC-PHAT, GCC-ML for time delay estimation delay estimation. These methods are based on spectral content of the speech signals that can be effected by noise and reverberation. Video tracking can be done using Kalman filter or Particle filter. Therefore Linear Prediction method is used for audio and video tracking. Linear prediction in source localisation use features related to excitation source information of speech which are less effected by noise. Hence by using this excitation source information, time delays are estimated and the results are compared with GCC PHAT method. The dataset obtained from [20] is used in video tracking a single moving object captured through stationary camera. Then for object detection, projection histogram is done followed by linear prediction for tracking and the corresponding results are compared with Kalman filter method.