Real-time Tracking Algorithm For Feature Compensation

Computer vision has always been a hot research field.In recent years,thanks to the rapid development of artificial intelligence,great breakthroughs have been made in this field,especially in the fields of image recognition and target detection.However,different from image recognition and target detection,which are more mature and have been applied in many real-world scenarios,many excellent algorithms have been proposed in the target tracking field and achieved unsatisfactory development.However,to meet the needs of practical applications,there are still many problems to be solved.Since the video sequence used for target tracking has a time dimension,uncertain factors such as illumination change,object deformation,scale change,occlusion and field of view will be faced in this process,which has become a difficult point to overcome in the research target tracking algorithm.This paper analyzes the tracking algorithm based on correlation filtering and the tracking algorithm based on neural network,proposes a tracking algorithm based on feature compensation,and combines the two algorithms based on correlation filtering and neural network through a logical regression classifier,in order to solve the former of poor performance in complex scene,long time tracking easy pollution template and the latter speed too slow.Simple and fast features are used in simple video scenes,while higher-level semantic features of convolution,which are more excellent but slower,are used in complex video scenes.Firstly,a correlation filter model with color histogram and direction gradient histogram is established in this paper to improve the robustness and algorithm speed of the algorithm when the color and shape of the target are changed.Then,a neural network tracking model is established,and the convolutional neural network is used to extract high-level semantic features,so that the algorithm can adapt to complex target tracking scenes,thus improving the accuracy and robustness of the algorithm in locating targets.Finally,a simple classifier is constructed for model switching,in which the algorithm uses color and direction histogram features in simple scenes and high-level semantic features in complex scenes,so as to effectively improve the real-time performance,tracking accuracy and robustness of the algorithm.In addition,the problems existing in the training data of the classifier are analyzed,and reasonable and effective solutions are proposed to make the classification effect more excellent.