Object Tracking In H.264 Compressed Domain Based On Temporal-spatial Markov Model

Object tracking is a hot spot in the field of computer vision and artificial intelligence,which has important research value in video retrieval,video analysis and pattern recognition.According to different operating domain,tracking algorithms are classified into pixel domain algorithms and compressed domain algorithms.Pixel domain algorithms totally decode video into image frames,and then use pixel domain information for target tracking.Most pixel domain algorithms do not have direct target motion information,they search the region similar to object features in the next frame to locate targets and implement track,leading to high computational complexity and difficulties in real-time processing.Video compression process generates advantageous information for target tracking,such as motion vectors(MV)and Discrete cosine transform(DCT)coefficients.MV reflects the macro block displacement between adjacent frames.DCT coefficients reflect the image texture.Compressed domain algorithms can extract the MV and DCT coefficient through partially video decoding,which not only obtain the useful information,but also reduce the decoding time.This paper builds temporal-spatial Markov model to track objects in H.264 compressed domain,the main work is the following:1.A tracking algorithm based on pixel motion vector is proposed.The tracking algorithm based on block motion vector can achieve a certain degree of tracking,however,the tracking accuracy will be dramatically dropped with the increase of frame number.Under such condition,the pixel motion vector algorithm is put forward,the algorithm significantly improves the tracking accuracy and avoids the problem of the sharp drop,but the phenomenon still exists that tracking accuracy slowly declines.Through the contrast experiments of two algorithms,we get a deeper understanding of the temporal characteristic of the MV and its defects on the track.2.The pixel motion vector algorithm is taken as the MV temporal characteristic,which is combined with MV spatial characteristic in the framework of Markov random field.By maximizing a posteriori probability,every macroblock is marked as background or object,then objects are located to achieve tracking from frame to frame.Firstly,the algorithm preprocesses MV,then establishes temporal-spatial Markov model for tracking problem and utilizes the Bayesian framework to boil it down to the following: block label category depends on weighted sum of the temporal continuous degree,spatial consistency and compactness.Among them,the temporal continuous degree is computed by the pixel motion vector algorithm.Comparative test shows that the proposed algorithm can improve the tracking accuracy.