Video Human Tracking Algorithm Based On SNAKE Models

The human-oriented motion tracking technology which has been used widely in intelligentvideo surveillance field is one of the important topcis in recent years. The process of trackingbased on common human target detection and tracking method is divided into imagepreprocessing, motion prospect area segmentation, body targeting and tracking.Active contour model get a joint solution about the segmentation or extraction issueswhich include the initial estimate,image data, target contour and constraint based on priorknowledge etc. and break up the limitations of information transmission direction.In theprocess of tracking tasks, an effective solution is proposed about how to selecte initial contourautomatically,decrease the high time-consuming of calculate GVF and solve unstable trackingproblems. In the process of tracking human targets, the direction and speed of movement of thehuman is unknown. When the target’s movement is faster, using GVF-Snake model is hard tokeep up with the speed of change in the target position,and then loss the target.In order to solve the problem that GVF-Snake model can not automatically obtain the initialcontour, we propose a way to automatically obtain a stable initial contour by detection method.It is necessary to get complete enough edge information with detection to obtain more accurateinitial contour line. We design an adaptive detection algorithm based on combination ofimproved background subtraction and three frame difference by comparing the advantages anddisadvantages of the traditional classic detection method. Our detection method can achieve astable outlook in the region containing the moving human target, and then get the body target byde-noising and positioning. Finally, we can get the GVF-Snake initial conour line we need byextracting and discreting goal outline.Furthermore, in order to solve motion loss problem when the GVF-Snake mode is used totrack fast moving target, we utilize the Kalman filter to predict motion trajectory with theobtained centroid point by GVF-Snake convergence as the movement characteristics of thehunman body target. The current estimate of the Kalman filter is estimated based on the currentframe measured values and the last frame estimated value including the movement informationat last time to reduce the probability of motion loss. The combination of Kalman filter andGVF-Snake can enabe the initial contour closer to the actual profile. We take use of the canoearea to improve the calculating speed because of the slow calculating the global gradient vectorflow field with the GVF-Snake.Finally, we do some simulation experiments for the detection and tracking algorithm. Onthe one hand, for the shaking camera leading to a large area of interference, we compare theadaptive difference method and the background subtraction method. The experimental resultsshow that adaptive differential method can detect the motion stably. On the other hand, we do the comparative experiments between the combination method of frame difference andGVF-Snake and our tracking method. We use frame difference alone to extract less targetinformation and more interference, and the lower quality of the initial contour extraction is ourfocus. Specially, without limitative sub-regions and forecasts of the Kalman filter, the contourlines will gradually be away from the target boundary. Experimental results show that ouralgorithm can make the contour line close to the real contour edge, and achieve good trackingresults and better real-time performance.