Research Of Visual Target Tracking Algorithm Based On Locality Sensitive Histogram

Vision target tracking can provide technological support for advanced tasks of computer vision(e.g., image understanding, behavior analysis, etc.), so it is one of the most active research fields in computer vision, widely used in augmented reality, intelligent video surveillance, object recognition, medical diagnosis, vision navigation, space surveillance, and other fields. At the same time, in recent years, the wide application of video sensor, test and measurement technology, and the research on multi-sensor fusion in the instrument field, have hastened the requirements to the automatic process of output data from video sensor, which also promotes the research on vision target tracking technology.Although many arithmetics have been put forward to resolve the video tracking problem in different application situations after a long time of deep research,there are still many difficulties when applied in actual environments. For example,the target appearance will change according to background, illumination change and sheltering by other objects. Meanwhile, unpredictable complicated changes such as shape change, rotation, scaling and motion blurring may happen to the target in the video during motor progress. All the above make it a great challenge to construct a vision target tracking algorithm with universality, precise and stable robustness in actual environments.The technique proposed in this dissertation improves the tracking effect of single visual target in complex environment by improving the target appearance modeling in visual tracking algorithm. Subsequently, on the premise of reducing no tracking accuracy, the tracking speed has been increased by reducing marked radius and Mean Shift preposition. This dissertation has completed the following tasks:On the base of researching Mean Shift visual target tracking algorithm and Haar's characteristics, the distribution of Haar eigenvalues in various quantizedintervals is changed through regularizing and quantifying the Haar eigenvalues.The method that Haar features and color features jointly build the target appearance model has been proposed by making further use of HSI color space in which brightness and colors are separately indicated. This improves the tracking effect of Mean Shift visual target tracking algorithm when the target is blocked by analogue and illumination varies.On the foundation of researching visual target tracking algorithm based on Locality Sensitive Histogram, firstly, the calculation defects of Locality Sensitive Histogram fast algorithm has been described; secondly, a new Locality Sensitive Histogram fast algorithm has been obtained by introducing L2 norm to the measurement of the spacing among pixels in ?45? direction of the original fast algorithm, thus amending the calculation defects in that direction; thirdly, the method of adaptive generation of ELSH suboptimal weight coefficient ELSH has been obtained by taking advantage of the distance metric difference among pixels of L2 norm and L1 norm in ?45? direction; finally, through introducing Bhattacharyya coefficient to determine the templates similarity, the deficiency of determining templates similarity by comparing specific position has been amended. Ultimately, visual target tracking algorithm based on 8 neighboring Locality Sensitive Histogram has been obtained, and the tacking effect of varied brightness target in this algorithm is superior to that of the original visual target tracking algorithm.The tracking time of visual target tracking algorithm based on 8 neighboring Locality Sensitive Histogram has been reduced by shortening the marked radius of 8 neighboring Locality Sensitive Histogram. The accompanying similarity decline among templates has been retarded by adopting Mean Shift preposition.The target template in Mean Shift is updated by using the updated modules in visual target tracking algorithm based on Locality Sensitive Histogram. Finally,the visual target tracking algorithm based on preposition 8 neighboring Locality Sensitive Histogram has been obtained. On the premise of reducing no tracking effect, the tracking time of visual target tracking algorithm based on preposition 8neighboring Locality Sensitive Histogram has been effectively reduced.The test on multi-group pointed video data shows that the tracking algorithm put forward in this dissertation has good performance on handling illuminationchange, sheltering, motion blurring, and similarity background.