Tracking Method Based On Deep Learning Feature Extraction

Target tracking is a video image processing method that uses a given target in the first frame as a template,searches for the target position in subsequent consecutive frames,and determines the motion state of the target.The key to target tracking is to accurately distinguish the foreground and background in the complex realistic environment of each frame,and to track the target stably for a long time.Target tracking has broad application prospects in real life,and there is still no method to track targets stably in all environments.Therefore,the research on target tracking is very meaningful,and it is urgent for more scholars to explore and study.In target tracking,deep learning methods are usually used for feature extraction.Excellent features help trackers to better distinguish foreground and background information in complex environments.Currently,the target tracking deep feature extraction method mainly uses Siamese network.The template and the search image branch share the same Siamese network,and the state of the target in the current frame is obtained according to the relevant feature map of the template and the search area.However,the existing target tracking methods for deep learning feature extraction are more focused on the detection of the target state in each frame and the effect of offline training,ignoring the relationship of tracking in time series and the role of online update.Therefore,on the basis of in-depth study of the target tracking method based on deep learning feature extraction and careful analysis of the shortcomings of the target tracking method of deep learning feature extraction at this stage,the Siamese network is used for feature extraction.This paper use Siamese network to do feature extraction and correct Siamese tracking results using Kalman filtering in traditional methods;on the basis of structural sparsity,propose to use attention mechanism to improve tracking performance of structural sparse representation methods.The main content of this article includes the following two aspects:(1)A target tracking method based on adaptive structured sparse representation and attention is proposed to deal with the problems of motion blur,partial occlusion and fast motion in target tracking.In the framework of particle filtering,the attention mechanism is used to improve the performance of high-quality templates.The sparseness of the structure is used to establish a sparse model between the candidate target set and the local patches of the target template.Combine the sparse residual method to reduce reconstruction errors.After the model is optimized and solved,the particle with the highest similarity is selected as the prediction target.Choose the most suitable scale according to the multi-scale factor method.(2)Use a Siamese network with segmentation branches to effectively segment the target object in the frame.Combining segmentation and tracking means that the tracking is completed according to the segmentation result,and the video object segmentation speed is optimized according to the tracking result.The lack of an effective online adjustment strategy makes the algorithm perform poorly in the face of occlusion problems.Aiming at the above-mentioned problems,a moving target tracking model based on time-space fusion combined with Kalman filtering is proposed to alleviate the occlusion problem in the tracking process.Combined with the segmentation model,the appropriate model is selected through the score to predict or detect the current state of the target.The ellipse fitting strategy is used to evaluate the bounding box online.