Research On Target Tracking Algorithm Based On Deep Features

Visual target tracking is an important branch in the field of computer vision.Due to the development of science and technology and the needs of daily life,such as drone monitoring,autonomous driving,human-computer interaction,and intelligent transportation,in recent years it has been able to demonstrate its skills on the historical stage.In addition,the rapid rise of deep learning has made big data analysis possible.This revolution breaks through the shackles of traditional algorithms,and can effectively respond to unpredictable changes in the target in actual scenes,such as background occlusion,deformation,rotation,and lighting changes.In the field of tracking,accuracy and robustness are usually used to measure algorithm performance.Compared to traditional tracking algorithms,deep learning-based trackers can greatly improve the robustness of the algorithm with its excellent ability to reconstruct data features.The feature has no deformation for translation and scale,and the spatial resolution is low,and the target cannot be accurately located.Based on this,this paper compares and analyzes traditional algorithms and deep learning algorithms.The main work is as follows:(1)Comprehensively expound the research status of traditional algorithms and deep learning algorithms in the field of target tracking,and point out the advantages and disadvantages of each method.First of all,traditional target tracking algorithms are mainly based on manual features for motion modeling.The calculation is small and fast,but the model is not robust.Convolutional features extracted by self-learning based on deep learning algorithms have rotation invariance and translation invariance,and are more robust.Secondly,the existing deep learning algorithms are summarized according to the network structure.(2)Aiming at the problem that traditional tracking algorithms have poor anti-occlusion ability and robustness in complex scenes,a moving target tracking algorithm based on adaptive fusion of deep features is proposed.Considering the strong robustness of deep features and the advantages of high precision of shallow features,this paper first uses sparse autoencoders to construct deep sparse features to extract targetfeatures,and then uses the sparse auto-encoder to extract the target features,and then uses the correlation information between adjacent frames and the tracking confidence to the depth features.Adaptive fusion with texture information to improve tracker performance.In order to improve the robustness of the tracking algorithm while suppressing tracking drift,when the confidence level is lower than the set threshold,an improved SURF(Speeded Up Robust Features)algorithm is introduced to relocate the target.Experimental results show that: compared with mainstream tracking algorithms,the tracking accuracy of the proposed algorithm is higher than that of the contrast method,and it has good robustness in occluded scenes,and can effectively suppress tracking drift.(3)Aiming at the problem of poor accuracy of deep learning-type tracking algorithms,a high-resolution network is introduced,and a multi-layer feature fusion structure for cross-correlation operations is proposed,which helps the tracker learn from multiple resolutions.Of the richer features in the prediction similarity map.Enhance high-resolution representations by maintaining high resolution during forward propagation while performing repetitive multi-scale fusion,with low-resolution representations of the same depth and similar levels,resulting in high-resolution representations with more lossless scale information.