Video Tracking Based On Locality Sensitive Histograms

Object tracking is an important research area in machine vision, which plays a significant role in the application field such as monitoring and control system, military defense construction, motion analysis, behavior recognition, human-computer interaction and so on. Therefore, research on object tracking has its practical significance. In the field of video tracking, a key factor that impacts the performance of the tracking algorithm is how to represent a target, which means the selection of appearance model for the target representation. Recently, two target representation methods based on layers had been proposed, which are distribution field representation and locality sensitive histograms representation. Two improved algorithms was proposed based on locality sensitive histograms representation. The main conclusions of the study are as follows:(1) A video tracking based on distribution field of locality sensitive histograms was proposed. For the similarity of locality sensitive histogram and distribution field model, a new distribution field based on locality sensitive histogram to represent target is put forward. Compared with the original distribution field algorithm, the new algorithm overcomes the defects of original distribution field that is sensitive to illumination and parameter and assigns larger weight to the layer which the pixel is in, enhances the robustness of the new algorithm and improve the performance of it. This algorithm uses L1 distance to measure the similarity of the target candidate block and the current block. The experimental results of representative videos show the new algorithm compared with the multiple instance learning tracking algorithm, local sensitive histogram tracking algorithm, the distribution field tracking algorithm and tracking by exploiting the circulant structure with kernels gets the best result by a fixed set of parameters. The average tracking accurate rate is increased by 23.55%, 10.33%, 12.81%, 3.18% and the average tracking error is reduced by 8.06, 3.69, 4.41, 3.13(in pixels). The average tracking speed is 10.72 frames per second.(2) A video tracking based on distribution field of multi-channel locality sensitive histograms and principal component analysis is proposed. At first, this algorithm gets local sensitive histograms about the three channels of color images, which makes up a three dimensional matrix. Then, the principal component analysis was applied on the three dimensional matrix. Each layer in the matrix is stretched to a vector, and is combined together to form a two-dimensional matrix used for representing an object. The correlation coefficient is used to measure the similarity of the target candidate block and the current target block to find the target location. Because of the full use of the color information of the target, so the algorithm has achieved good results. Compared with the four methods including distribution field tracking algorithm, local sensitive histogram tracking algorithm, tracking algorithm based on locality sensitive histogram distribution field and tracking by exploiting the circulant structure with kernels, the average tracking accuracy of the new method is increased by 12.48%, 10.49%, 2.07%, 8.56%. The average tracking deviation is reduced by 4.59, 2.32, 0.29, 3.54 and the average tracking speed is 8.17 frames per second.It can be concluded through the experiment comparison, the two proposed tracking algorithms have good adaptability for the target rotation, occlusion, illumination change, similar backgrounds and other complex cases. Compared with the other four representative algorithms, the two proposed tracking algorithms achieved better tracking results.