Research On Single Target Tracking Algorithm Based On Deep Learning In Complex Scenes

Tracking of video targets is a research hotspot in the field of computer vision.Combined with target segmentation,it can achieve the separation of research target and background and the follow-up of target motion trajectory.This technology is widely used in video traffic detection,driverless,industrial automation,medical diagnosis,military reconnaissance and other fields.The existing video target tracking and segmentation algorithm works well under normal scenes,but the segmentation is inaccurate in complex scenes such as light changes,target scale changes,target occlusion,target deformation,fast target movement,and background clutter which can causes tracking drift and other conditions.Aiming at the above problems,this paper proposes an improved target tracking algorithm to solve the single target tracking drift problem in complex scenes and improve the accuracy and robustness of tracking.Mainly in a complex scenario,using a single camera to track a single target.The main research contents and innovations of this paper are as follows:(1)An improved target segmentation algorithm is proposed to process the initial frame image,not only to determine the position of the target,but also to determine the accurate contour of the target,to achieve unsupervised learning.We do not need to manually process the initial fram.Using semantic segmentation based on convolutional neural networks,images are processed from the pixel level to achieve more accurate and robust segmentation,which lays a solid foundation for subsequent tracking.(2)An improved target tracking algorithm based on neural network and correlation filtering is proposed.Combining the high accuracy of the target tracking algorithm based on neural network with the rapidity of the target tracking algorithm based on correlation filtering,the three aspects of template update,target size and image processing are optimized to achieve the unification of accuracy and speed.(3)An improved correlation filter in the target tracking algorithm is proposed.The classic Kernelized Correction Filters(KCF)have poor tracking performance when the target scale changes,rigid deformation,and rapid motion occur.Based on the improvement of this algorithm,a kernel-based ridge regression model is proposed.The neural network is used to solve the model,and the shallow and deep features are extracted to improve the accuracy ofclassification and image localization.(4)Introducing spatial regularization kernel into a complete convolutional neural network.By applying a spatial constraint to the kernel filter,each output channel of the convolutional layer is forced to respond to a particular local region,and the distance transform pool layer is utilized to determine the validity of the convolutional layer output.Compared with the neural network-based tracking algorithm MDNET,the speed of the algorithm is improved by 6 times.Through the above four aspects of research and experimental verification with the character as the target,the accuracy is improved by 2.6% and the success rate is increased by9.6%.It is shown that the target tracking algorithm based on neural network and correlation filtering proposed in this paper can achieve the accuracy and robustness of target tracking in complex scenarios.