University of Houston

Change detection methods applied to images of the same scene captured by a fixed camera is well studied, but there are few work on change detection of scene in temporally separated videos taken from a mobile camera. This thesis provides a comprehensive quantitative comparison of metrics for detecting visual anomalies between two videos that are recorded along same path but at different times by a camera on a patrolling platform. The metrics used in this thesis are histogram based metrics, statistic based metrics and pixel differences based metrics. The metrics are tested for the detection of mobile and stationary anomalies between videos. The two videos are brought to spatial temporal alignment by a two step process. For each frame in the first video the closest matching frame form the second video is found manually and the matched pair of frames are registered using a feature based registration method. Intensity invariant Laws texture kernels are used to extract texture energy measures from the images and nine different metrics are applied to generate a difference image sequence which is followed by thresholding and blob coloring algorithm to get a binary image sequence. The binary images are compared with the actual ground truth and the performance of each metric are presented for four videos taken in different environments.