On Video Target Tracking Algorithm Based On Structured Output SVM

With the rapid development of artificial intelligent,the visual target tracking technology has also made great progress.The tracked target is no longer limited to face and people,but applies to all moving objects.Stabilizing the tracking of objects by overcoming different complex scenes has become a key technical of current research.Based on the target tracking algorithm of structured output,the dual update strategy adaptive target tracking algorithm is designed.The thesis is divided into the following parts:The first part is image preprocessing.The second part is structured output tracking.The third part is the dual update strategy adaptive tracking.The fourth part is the specific scene tracking.The image preprocessing mainly performs feature extraction and sample classification on the objects in the image.The Haar feature and HOG feature used for image feature extraction are theoretically analyzed and calculated separately,and the acquired target images are classified into positive and negative samples using support vector machine.The structured output tracking part is the main frame of target tracking.The image features are classified by structured support vector machine,and the continuously moving and changing targets are tracked by using an online learning method.In the adaptive tracking part of the dual updating strategy,the similarity function is designed to similarly measure the eigenvalue of the target image sample.The 80%benchmark similarity threshold is set as a constraint for the online learning update positive sample,and the 95%high similarity threshold is set as the target quick search scheme.It can overcome the complex scenes such as large-scale or full occlusion to track the target stably.The specific scene tracking part is divided into two cases:specific object tracking and fixed background tracking.Target sample library is built for tracking specific objects,and samples of the sample library are trained by using offline learning.For the tracking of general targets in a fixed-light background,the region of interest is used to further narrow the scope of the search target.The experimental results show that the dual-update strategy adaptive visual target tracking algorithm designed in this thesis can effectively overcome the interference of complex scenes,especially the occlusion situation.The performance of the algorithm tracking target can be verified by parameters such as center position error.It can consistently track general goals and specific targets with real-time,accuracy and small error.