Research On Correlation Filtering And Tracking Algorithm Based On Convolutional Neural Network Multi-feature Fusion

Object tracking plays a vital role in many real-time visual applications such as video surveillance systems,unmanned driving,human-computer interaction,video communication,traffic safety and medical diagnosis.In recent years,the target tracking algorithm based on correlation filtering has attracted the attention of scholars with its amazing speed and good accuracy,and has continuously achieved brilliant results.However,target tracking is a very challenging task.For targets in motion,the actual moving scene is very complicated,such as lighting changes,background clutter,or the appearance of the moving target itself is deformed,such as Rotate,be blocked,etc.In the actual research process,how to design an algorithm that is robust to various situations is a problem we need to pay attention to.In response to these problems,this paper uses a correlation filtering algorithm that combines multiple features and combines depth features to improve the performance of the traditional correlation filtering target tracking algorithm.The main work and innovations of this paper are as follows:A target tracking algorithm based on contextual filtering based on fusion features is used.When implementing target tracking through the correlation filtering algorithm,since the traditional correlation filtering algorithm often uses a cosine window to suppress the edge effect or expand the search range when the target is located(generally it will be doubled),what can be used in the target tracking algorithm The background information will be greatly reduced,so that the target drift phenomenon will often occur during target tracking.Based on the shortcomings of the above traditional related filtering framework,a context-based target tracking algorithm is adopted,and the use of HOG features and CN features is proposed.The fusion is used as the input feature,and finally the APCE occlusion mechanism is designed to judge whether it follows the new model to enhance the accuracy and robustness of the target tracking algorithm,and experiments show that the use of this tracking algorithm is stronger than the traditional several filtering algorithms.Robustness.The HOG feature is calculated on the local cell of the picture,and the geometric and optical changes of the image are very good for it.However,the HOG feature is not easy to detect when the target is blocked,fast-moved,or rotated.And it is very sensitive to noise.The CN feature hardly depends on the size,direction and viewing angle of the image itself,and is not affected by rotation and position changes.Therefore,it has high robustness,but it cannot describe the local features of the target well.The high-level convolutional features of convolutional neural networks have rich semantic information,can solve non-rigid deformation,occlusion and other problems,and can discriminate targets between classes,which is robust to target shape changes.Therefore,we multichannel fusion of the 31-channel HOG feature,11-channel color feature and convolution feature.In the training phase,the reliable weights are calculated according to the feature response values;in the positioning phase,the feature detection response values are reliably weighted to obtain the target position;in addition,for the model pollution caused by occlusion and other problems,an APCE occlusion mechanism is designed to Determine the model update.Finally,the data set(OTB-100)is used to perform qualitative and quantitative experimental analysis on the video sequence sequence with multiple challenge attribute annotations,and the performance comparison with various related filter tracking algorithms.The experimental results show that the proposed algorithm is the target It has good robustness and accuracy when rapid movement,deformation,occlusion,line of sight,etc.occur.