Research On Intelligent Target Tracking Algorithm Based On Multi-feature Fusion

With the rise of the third wave of artificial intelligence,with the help of the continuous development of big data technology and the strong problem-solving ability of deep network,deep learning continues to break through the challenges brought by computer vision.As an important part of computer vision,target tracking is widely used in all aspects of life,such as road monitoring,behavior analysis,and so on.In the past decades,new theories and algorithms have been emerging in target tracking at home and abroad,and achieved the corresponding research results.But at present,the mainstream target tracking algorithms mainly face the following problems:Firstly,Most of the traditional target tracking algorithms are extremely unstable under the influence of the change of the target itself and external factors;Secondly,When the target is more than one or the whole image is relatively small,the traditional algorithm and shallow machine learning algorithm cannot track the target with high precision;Thirdly,in fact,target tracking has a high demand for real-time and robustness.Most of the traditional target tracking algorithms are not robust enough,and the target tracking algorithm based on deep learning does not achieve a good balance between accuracy and speed.This paper puts forward effective solutions to the above problems.The details are as follows:(1)Aiming at the problems of poor accuracy and real-time performance of traditional target tracking algorithm under the interference of external light changes,target occlusion,motion blur,a multi-feature fusion target tracking algorithm(SDDAE-MultiFF)based on the depth noise reduction automatic encoder is studied and proposed in this paper.In this method,the image enhancement algorithm is improved and applied to the preprocessing of self-collected image,and the quality of training set data and test set data is effectively improved;An adaptive threshold texture feature with invariable scale is proposed,then the feature and color feature of the target are extracted,weighted fusion is carried out,depth neural network is used for learning,and an automatic coding network of multi feature depth noise reduction is constructed;Then the confidence score is obtained through the logic regression classifier,and further effectively differentiate goals and backgrounds.In the end,Particle filter framework is used to track the target.Compared with the traditional three algorithms and six popular algorithms,this method can not only track the video with interference problems more accurately,but also meet the requirements of real-time processing.(2)In order to solve the problems of multi-target and weak small target tracking,a target tracking method(YO3-FKCF)which combines yolov3 with fast kernel correlation filtering is processed in this paper.In this method,the image mosaic method is used to expand the self-collected data set,and through the weighted integration of Mahalanobis distance and minimum cosine distance,and the K-Neighborhood search method,the yolov3 network model is improved;Then the yolov3 network model is trained by the training method of migration learning to integrate the target motion information and apparent information.In so doing,the practical model combining the self-provided training set can be obtained,and the accuracy of target detection in multi-target scenes and small and weak targets can be effectively improved.Finally,construct fast kernel correlation filters and combine them for target tracking.The experimental results show that the target detection part improves the detection accuracy of multi-target and weak small target scene after improvement,and the target tracking method after combination also has high accuracy,strong robustness and suitable real-time performance in the corresponding scene tracking process.