Fast Tracking Based On Local Histogram Of Oriented Gradient

Object tracking is an important branch of computer vision and a current research hotspot because of its application.Currently,object tracking has been applied to many areas,such as: video surveillance,intelligent monitoring,unmanned aerial vehicles and face recognition and so on.Object tracking refers to a series of operations on detection,extraction,identification and tracking of a particular moving object in a sequence of digital images,and obtaining various motion measurement parameters of the specified moving object,such as position,motion trajectories,etc.,in order to process and analyse the tracking data,so as to understand the behavior of the specified moving object.At present,there are many different categories of tracking algorithm.The tracking algorithm based on the detection is most popular.There are still many problems that need to be solved about tracking,for example,the faster tracking speed needed by actual applications,occlusion about object.Based on these existing problems,this paper has obtained research results,as followings:Firstly,the histogram of oriented gradient is chosen as the feature of tracking algorithm.Under normal circumstances,the sampled area,which extracts feature,includes the entire digital image.But a large number of data sampled is invalid,whose relationship with the object is not close.Therefore,this paper proposes a local sampling strategy for an image,which reduces the computational complexity of the algorithm by sampling a certain multiple rectangular region according to the predicted target position instead of the entire image.Secondly,in the processing of the software,the algorithm speed is reduced due to the processing of the loop because the feature matrix of the acquired histogram of orentied gradient is three-dimensional data.In this paper,the preprocessing of the feature matrix data is proposed,that is,the matrix of image feature is reduced about dimension,which reduces the computational complexity of the tracking algorithm and improves the algorithm speed.The main operation is to convert the three-dimensional feature matrix data into two-dimensional data,which reduces the number of loops of the operation to improve the tracking speed,at the same time,guarantees the spatial relationship between the data space.Thirdly,the dual detection is proposed to improve the detection efficiency about the object occlusion and reduce the loss of the object target.The so-called dual detection refers to the second detection of the occluded image when the occlusion has occurred in former image.When the occlusion occurs for the first time in the image sequence,a set of alternate feature information is stored while the tracking detector updates the characteristics of the occlusion region.For the next image sequence,when an occlusion is detected by the tracking detector,then the backup information is used for second detection,to confirm whether the occlusion or changed location of the object target.The dual detection is to prevent the loss of the target caused by the detection errors about target occlusion.Next,through standard test platform,the proposed algorithm and a variety of popular tracking algorithms are experimented comparatively.And the experiments show that the algorithm proposed in this paper is much faster than the contrast algorithms,and some progresses have been made in the detection efficiency of occlusion in some sequences.Finally,the author summarizes this paper.According to the content of this paper and the application of practical needs,the author proposed a better direction.