Visual tracking in unstabilized real time videos using SURF

Abstract

Auto target detection and tracking has always been a challenging task for the researchers. The problem becomes worse, if the input video to tracking algorithm is itself unstabilized. This complex problem of auto target detection and tracking in an unstabilized video is not well addressed and solved in the literature. An algorithm is developed and presented in this paper to auto-stabilize the video, auto detect the target of interest followed by tracking. Global and local motion vectors are computed in the same algorithm simultaneously for frame stabilization and target tracking respectively. Speeded Up Robust Features (SURF) are used innovatively to compute Global Motion Vector (GMV) and Local Motion Vector (LMV) separately. The algorithm auto-detects the unstabilized parameters if any and stabilizes the video by moving the frame in opposite GMV direction with anti rotation. First four frames are used for detecting and training the target of interest, then only two very recent frames are used for auto-detection of target, hence algorithm is very fast and suitable for online real time applications. The algorithm has been experimentally verified and outperforms with bench marked visual tracking algorithms.